Datasets:

jamescalam

/

ai-arxiv-chunked

like 32

2106.09667

16

that any image xwith a particular backdoor pattern bd(denotedxbd) will be classified incorrectly. The only change we make to turn our poisoning attack into a backdoor attack is instead of always using the same image x0that is paired with various captions, we use different images xibdfor each poison sample. Specifically, we define P=f(xibd;c) :c2caption set; xi2X subsetg. Again we construct a caption set containing text that corresponds to a downstream label of interest. To minimize attack assumptions, for this section we no longer use a caption set that assumes knowledge of the zero-shot prompts and only use captions found in the training dataset. 1While this is without loss of generality—and the adversary may indeed have wanted to cause gto be modified—we have specified the attack objective in advance. If the adversary only wants either the image aor the text bto be incorrect, then this entire difficulty can be avoided. 4 Published as a conference paper at ICLR 2022 100101102 Number of Poisoned Samples0.00.20.40.60.81.0Probability Attack SucceededCC3M zero-shot

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

17

100101102 Number of Poisoned Samples0.00.20.40.60.81.0Probability Attack SucceededCC3M zero-shot CC3M linear probe YFCC zero-shot 150300 1500 Number of Poisoned Samples0.00.20.40.60.8Probability Attack Succeededzero-shot linear probe Figure 2: Left: Poisoning attack success rate on Conceptual Captions-3M and YFCC when inserting between 1 and 512 poisoned examples (datasets with 3 million and 15 million images respectively). Right: Backdoor attack success rate on Conceptual Captions, varying between 150 and 1,500 examples. The shaded region corresponds to one standard deviation of variance. 4 E VALUATION We now investigate to what extent our poisoning and backdooring attacks are a realistic threat on multimodal contrastively trained models. 4.1 E XPERIMENTAL METHODOLOGY We demonstrate the efficacy of our attack on two datasets: the 3million example Conceptual Captions dataset (Sharma et al., 2018), and the 15million example YFCC Thomee et al. (2016) subset. Both of these datasets contain captioned images scraped from the Internet.

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

18

of these datasets contain captioned images scraped from the Internet. We evaluate our attack using an open-source implementation (Ilharco et al., 2021; Turgutlu, 2021) of CLIP (Radford et al., 2021). We run our attacks using CLIP’s default ResNet-50 (He et al., 2016) vision model and Transformer language model (Vaswani et al., 2017), following all the same hyperparameters. All our experiments use a batch size 1024, training across 8 V100 GPUs for 30epochs using a learning rate of :0002 training with Momentum SGD and weight decay of 0:02. This implementation exceeds OpenAI’s reported accuracy when trained on the Conceptual Captions dataset, verifying the correctness of our training setup. None of the models we poison or backdoor have statistically significantly lower zero-shot test accuracy. 4.2 P OISONING EVALUATION Figure 2 presents our main poisoning results, showing attack success rate as a function of the number of poisoned examples. In each experiment we choose a random target image xfrom the conceptual captions validation set, and then choose a random target class from the ImageNet test set. We then

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

19

of poisoned examples. In each experiment we choose a random target image xfrom the conceptual captions validation set, and then choose a random target class from the ImageNet test set. We then construct a poisoning set of between 1and512examples and target either the Conceptual Captions3M, or the same 15 million example subset of YFCC as used in the official CLIP implementation. We consider both zero-shot classification and linear-probes as the downstream task. In both cases we follow the same attack process outlined in Section 3.1. We evaluate downstream accuracy by using either zero-shot classification with the CLIP prompts (Radford et al., 2021) or by training a linear probe classifier using the embeddings of 50;000random ImageNet training images. To compute the attack success rate, we train 32 different models and measure the fraction of poisoned models for which f(x0) =y. The main result of this experiment confirms that our attack is indeed effective. Even by poisoning just three samples out of the 3 million examples in the conceptual captions dataset, we can fool the model into misclassifying targeted samples x0as one of 1000 different ImageNet class labels with 40% probability under zero-shot classification. In contrast,

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

20

different ImageNet class labels with 40% probability under zero-shot classification. In contrast, attacking semi-supervised learning requires a poisoning 0:1%ratio, a factor of 1000higher (Carlini, 2021). And despite being 5as large, poisoning a YFCC-trained classifier isn’t much harder than poisoning a CC-3M classifier (e.g., poisoning 15 of 15 million images succeeds 20% of the time). 5 Published as a conference paper at ICLR 2022 0.0 0.2 0.4 0.6 Similarity between f(xi) and f(xj)104 103 102 101 Probability c(xi)=c(xj) 0.2 0.0 0.2 0.4 0.6 0.8 1.0 Pairwise Cosine Similarity0246810Frequency (*104)(,2) curve Natural Data Histogram Backdoored Data Histogram Figure 3: Left: The similarity between two ImageNet validation examples xiandxjunder the embedding function fdirectly predicts the likelihood that the two images will have the same true label on the downstream task. Right: By poisoning 0:01% of a training dataset, we can backdoor

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

21

label on the downstream task. Right: By poisoning 0:01% of a training dataset, we can backdoor CLIP so that any two images with a trigger pattern applied will have a pairwise similarity of 0:78. This is five standard deviations about what we should expect, when comparing to the similartiy of natural, non-backdoored images that typically have a similarity of 0:1. 4.3 B ACKDOORING EVALUATION We now investigate the effectiveness of our backdooring attack. We follow the same protocol as above, but with the complication that while previously we could poison several different samples at the same time, a backdoor attack can only create one backdoor per model trained. Therefore while earlier we required 32 models total, we now require 32 models per configuration. We experiment with three different rates of poisoning ( 0:0005% ,0:01%, and 0:05%), since this requires ( 33212) 10;000GPU hours of compute. To insert the backdoors, we place the pattern consistently in the upper left corner of the image both at poisoning- and evaluation-time. We again find our attack to be effective even at these exceptionally low backdoor ratios: even at a 0:01% poison ratio (one in ten

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

22

effective even at these exceptionally low backdoor ratios: even at a 0:01% poison ratio (one in ten thousand samples), we reach a 50% attack success rate at backdooring zero-shot classifiers. Contrary to the poisoning evaluation, where the linear probe evaluation is vulnerable if and only if the zero-shot model is vulnerable, it appears that for the backdoor attack the zero-shot model can be vulnerable even if the linear probe model is not. Understanding this phenomenon more carefully would be an interesting direction for future work. 5 A BLATION STUDY Having seen that it is possible to poison and backoor contrastively trained models, it remains an interesting question to understand why it is possible. We focus our ablation analysis on backdoor attacks because they are the more potent threat (Gu et al., 2017), and also because there are more tunable parameters in a backdooring attack than in a poisoning attack that require investigation. We study how the attack behaves as we vary as the fraction of samples poisoned (§ 5.1.1), the patch size (§ 5.1.3) and the model and training data sizes (§ 5.1.2). 5.1 A STABLE METRIC :BACKDOOR Z -SCORE Before directly delving into performing significant new experiments, we consider the problem of

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

23

5.1 A STABLE METRIC :BACKDOOR Z -SCORE Before directly delving into performing significant new experiments, we consider the problem of designing a more stable metric to measure the efficacy of backdoor attacks. Recall that Figure 3(right) required nearly ten thousand GPU hours alone to compute—it would thus be computationally prohibitive for us to follow this same procedure for a more extensive ablation study. Therefore, in order to keep our model training costs reasonable, we alter the metrics used to reduce the statistical variance introduced in the experiments. Instead of reporting results as a function of 6 Published as a conference paper at ICLR 2022 75 150 300 600 1500 Number of Samples, Poisoning Consistently01234567Backdoor Z-ScoreTest on Consistent Patch Test on Random Patch 75 150 300 600 1500 Number of Samples, Poisoning Randomly01234567Backdoor Z-ScoreTest on Consistent Patch Test on Random Patch Figure 4: Attack success rate as a function of number of poisoned examples inserted in the 3 million sample training dataset (i.e., ranging from 0:0025% to0:05%). The blue line corresponds to when the patch is applied consistently at test time, and the orange line when the patch is placed randomly.

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

24

the patch is applied consistently at test time, and the orange line when the patch is placed randomly. Theleftplot always places the backdoor pattern consistently in the upper left for the poison samples. Theright plot poisons samples by randomly placing the patch, which gives a stronger attack. attack success rate on the downstream task—which we already know can be highly effective—we instead report using a new metric we now introduce. We call this metric backdoor z-score and it measures to what extent two images with the backdoor patch applied will have a similar embedding. Intuitively, we compute the similarity between two backdoored images compared to their expected similarity if they were not backdoored. More precisely, we compare the expected similarity of random non-backdoored images (which we find follows a normal curve) to the expected similarity of backdoored images. Definition 1 Thebackdoor z-score of a modelfwith backdoor bdon a datasetXis given by  Mean u2X;v2X hf(ubd);f(vbd)i Mean u2X;v2X hf(u);f(v)i
 Var u2X;v2X hf(u);f(v)i1 :

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

25

 Var u2X;v2X hf(u);f(v)i1 : In Figure 3(right) we observe that random images (the blue region) tend to have a pairwise cosine similarity near 0:1for this model: random images are general not similar to each other. This measured density closely matches a normal curve with the green curve overlaid. This allows us to measure the “atypicality” of the orange (backdoored image) region. Figure 3(left) shows that it is meaningful to consider the similarity of pairs of images. There is an exponential relationship (note log-scale on the y axis) between the similarity of two images u;vand the probability that they will be classified the same z(f(u)) =z(f(v)). Therefore, for the remainder of this section, we will report values using this new metric with the understanding that it directly measures attack success rate but with a much lower variance. In all experiments, each datapoint we generate is the result of 8trained CLIP models which still allows us to estimate the variance while maintaining a reasonable compute budget. 5.1.1 B ACKDOOR ATTACK SUCCESS RATE AS A FUNCTION OF POISONED FRACTION

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

26

maintaining a reasonable compute budget. 5.1.1 B ACKDOOR ATTACK SUCCESS RATE AS A FUNCTION OF POISONED FRACTION As a first experiment we repeat the earlier figure and investigate how the number of poisoned examples impacts the attack success rate. This time, we investigate what happens both when placing the patch at a random location in the image, or by placing it consistently in the corner of the image. Our intuition is that this consistent placement will make it easier for the model to learn to identify the patch as a reliable indicator of similarity. Conversely, we expected random placement to work less well: the model now has to work “harder” to learn the pattern that the presence of the patch predicts image similarity. We perform 80individual experiments of our backdoor attack. For each of 5different poisoning ratios (from 0:0025% to0:05%) and for the two different methods of either poisoning randomly or consistently, we run 8independent trials to establish statistical confidence. 7 Published as a conference paper at ICLR 2022 105106 Size of Training Dataset01234567Backdoor Z-Score75 poisoned samples 300 poisoned samples 0 5 10 15 20 25 30 Model Parameters (millions)01234567Backdoor Z-Score

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

27

300 poisoned samples 0 5 10 15 20 25 30 Model Parameters (millions)01234567Backdoor Z-Score Figure 5: Evaluating the scalability of our attack. Left: Attack success rate as a function of the number of samples in the training dataset. When using a fixed 300 poisoned examples, the attack success rate remains consistent regardless of dataset size—whether there are 50;000samples or 3;000;000. At a fixed 75 poisoned samples the attack success rate remains high until the dataset reaches a million samples (a poison ratio of <0:01%), but degrades at two and three million samples. Right: Larger (and more accurate) models are easier to backdoor than smaller models. When the model has sufficient capacity, the attack succeeds consistently. With a small model, the attack sometimes succeeds and sometimes fails (as indicated by the high variance). The results of this experiment are given in Figure 4. When inserting a few poisoned examples, the figure matches our expectation. For example, with 75poisoned examples ( 0:0025% of the dataset), a consistently-placed backdoor patch results in z-score of 2:5when evaluated on patches that are also placed consistently. (When the patches are placed randomly at test time, the z-score degrades as

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

28

also placed consistently. (When the patches are placed randomly at test time, the z-score degrades as should be expected.) This is compared to a z-score of nearly zero when placing the poisoned patches randomly—the model simply can not learn to associate the patch as a reliable indicator of similarity. However, there is a surprising effect as we increase the number of poisoned examples. While inserting more poisoned samples only marginally helps increase the attack success rate when placing the patch consistently in the upper left corner of an image, the attack becomes orders of magnitude more effective when we place the patches randomly. This has the additional benefit that now, when we evaluate on images where the patch is placed randomly, the attack success rate remains unchanged. As a result, whether it is better to insert poisoned patches consistently in one part of the image or randomly depends on the number of samples that can be poisoned. When poisoning less than 0:01% of the dataset (i.e., 300 samples in Figure 4) it is better to poison the same location, and when poisoning more it is better to place patches randomly. 5.1.2 B ACKDOOR ATTACK SUCCESS RATE AS A FUNCTION OF MODEL AND DATA SCALE This ablation section studies a large (29 million parameter) model trained on a large (three million

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

29

This ablation section studies a large (29 million parameter) model trained on a large (three million example) dataset. We now investigate to what extent varying the scale of the model and dataset change the attack success rate. Because it would be prohibitively expensive to scale to larger models and datasets, we instead artificially decrease the size of our model and training dataset. Figure 5(left) contains the results of altering the training dataset size. Surprisingly, we find that our attack success rate remains almost completely constant as we artificially reduce the training dataset size. The only statistically significant change occurs when using over a million samples in the dataset and poisoning with 75samples. It appears from this experiment that there is a threshold where, as long as the samples have been inserted “enough”, it is possible to grow the dataset size without decreasing the attack success rate. Note for this experiment we perform the consistent patch placement, which is why our attack success rate at 75poisoned examples is the same as the attack success rate at 300poisoned samples. Figure 5(right) gives the results of varying the model size. Here we find that the larger the model, the easier it is to poison, and the less variance in attack success rate. For example, while a 1million

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

30

the easier it is to poison, and the less variance in attack success rate. For example, while a 1million parameter model is never successfully backdoored, a 5million parameter model sometimes has a 8 Published as a conference paper at ICLR 2022 z-score of 5:4and sometimes a z-score of 0:3. As we grow the model to 30million parameters, not only does the average attack success rate increase, but the variance decreases to the point that for a 30 million parameter model, the z-score is always between 5:1and5:9 5.1.3 B ACKDOOR ATTACK SUCCESS RATE AS A FUNCTION OF PATCH SIZE 10x10 20x20 30x30 Patch Size012345Backdoor Z-Score Figure 6: Attack success rate as a function of backdoor patch size, poisoning 0:0025% of the dataset. As the patch increases to 44the attack begins to succeed. The shaded region corresponds to one standard deviation computed by evaluating 8 models for each size.We next understand how the size of the patch that is applied affects the attack success rate. Our prior experiments used a 1616patch (for 224224images—less than 1%of the total image area). We find

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

31

1616patch (for 224224images—less than 1%of the total image area). We find that while small 22patches can not effectively poison a model, once the patch size becomes 44the attack already succeeds (see Figure 6). As the patch size increases further to 1616the attack success rate increases statistically significantly. Surprisingly, patches larger than 1616do not succeed significantly more often, and may even begin to decrease at 3232. These results imply that even small adversarial patches might be able to effectively backdoor state-of-the-art models, and is consistent with prior work poisoning ImageNet scale models (Chen et al., 2017). 6 C ONCLUSION Machine learning has traditionally been used in settings with a carefully constructed problem setup (e.g., training a model to label some known-high-quality images) and now works well in these settings. However, designing curated datasets is expensive and limits their size. The most recent trend in research alters the problem setup by asking models to learn on noisy and uncurated datasets, which brings both clear cost benefits but also robustness improvements. In our paper we demonstrate that training on this these unfiltered datasets, while now possible,

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

32

brings both clear cost benefits but also robustness improvements. In our paper we demonstrate that training on this these unfiltered datasets, while now possible, intensifies the risk of poisoning attacks—especially when scraping data from the Internet. Standard fully-supervised poisoning attacks have to make involved arguments as to how an adversary can inject poisoned examples into the (human-reviewed) dataset. Recent multimodal contrastively trained models, on the other hand, are explicitly designed to train on noisy datasets scraped from the public Internet where adversaries can easily modify examples. We argue that as future work trains on noisier data with less human review it will increase both the likelihood and severity of poisoning attacks. Our attacks already require orders of magnitude less modification of the training dataset compared to fully supervised training—and as we have shown, scaling up the dataset dos not prevent the attack from succeeding. The existence of these attacks motivates future defense research. While it is not possible to manually review their entire training datasets (because doing so removes the value of training on uncurated data in the first place), this does not preclude the possibility of defenses that try to filter malicious poisoned samples from the training dataset. For example, in the semi-supervised case it is possible

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

33

poisoned samples from the training dataset. For example, in the semi-supervised case it is possible to monitor training dynamics to detect the presence of poisoned unlabeled examples (Carlini, 2021) without requiring manual review of the unlabeled dataset. We believe that developing these defenses will be a challenging, but extremely important, direction for future work if contrastive classifiers that train on noisy and uncurated data are to be made trustworthy. Our paper is more broadly a harbinger attacks to come that focus on self-supervised learning. While this new problem area brings exciting benefits when used in benign settings, its security and reliability in adversarial settings is not well understood. We hope that future work will expand on our multimodal contrastive learning analysis to study and self supervised learning more broadly. 9 Published as a conference paper at ICLR 2022 ACKNOWLEDGEMENTS We are grateful to Kihyuk Sohn and the anonymous reviewers for feedback on drafts of this paper. ETHICS STATEMENT Our paper develops a practical attack on current multimodal contrastively trained classifiers. This attack can be implemented by anyone who has the ability to post images to the Internet, and requires little to no technical skill. While this might make our paper seem harmful, we believe the benefits of

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

34

little to no technical skill. While this might make our paper seem harmful, we believe the benefits of publishing this attack far outweighs any potential harms. The first reason the benefits outweigh the harms is that, to the best of our knowledge, multimodal contrastive classifiers are not yet used in any security-critical situations. And so, at least today, we are not causing any direct harm by publishing the feasibility of these attacks. Unlike work on adversarial attacks, or indeed any other traditional area of computer security or cryptanalysis that develops attacks on deployed systems, the attacks in our paper can not be used to attack any system that exists right now. Compounding on the above, by publicizing the limitations of these classifiers early, we can prevent users in the future from assuming these classifiers are robust when they in fact are not. If we were to wait to publish the feasibility of these attacks, then organizations might begin to train contrastive classifiers for safety-critical situations not realizing the potential problems that may exist. Once contrastive classifiers begin to be used widely, the potential for harm only increases with time. Finally, by describing the feasibility of these attacks now, we maximize the time available for the

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

35

Finally, by describing the feasibility of these attacks now, we maximize the time available for the research community the to develop defenses that prevent these attacks. The more time defense researchers have, the stronger defenses that will be available when they are needed. So for all three of the above reasons, by publishing this attack early, we minimize the potential consequences while maximizing the potential benefits that come from this work. This line of reasoning is not new to us, REPRODUCIBILITY STATEMENT There are two aspects of reproducibility to consider for this paper. The first is if it is possible to reproduce our paper. Here the answer is yes, and indeed it is fairly easy: our attacks only require running existing open-source CLIP training tools out-of-the-box on a slightly modified training dataset (i.e., those with poisoned samples). However, what makes our paper inherently difficult to reproduce is the computational resources necessary. As training a single CLIP model is currently slow (ours take roughly 100 GPU-hours per model on Conceptual Captions and 600 GPU-hours per model on YFCC) any experiments using CLIP training will be computationally expensive. Fortunately, here, we believe that because we have already comprehensively evaluated the attack across various

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

36

on YFCC) any experiments using CLIP training will be computationally expensive. Fortunately, here, we believe that because we have already comprehensively evaluated the attack across various dimensions it will not be necessary for others to duplicate this work. Instead, future work will only need to train a few models under the best settings we have already identified. REFERENCES Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644 , 2016. Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910 , 2019. Marco Barreno, Blaine Nelson, Russell Sears, Anthony D. Joseph, and J. D. Tygar. Can machine learning be secure? In Proceedings of the 2006 ACM Symposium on Information, Computer and Communications Security , ASIACCS ’06, pp. 16–25, New York, NY , USA, 2006. Association for Computing Machinery. ISBN 1595932720. doi: 10.1145/1128817.1128824. URL https: //doi.org/10.1145/1128817.1128824 .

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

37

//doi.org/10.1145/1128817.1128824 . Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. InInternational Conference on Machine Learning , 2012. 10 Published as a conference paper at ICLR 2022 Battista Biggio, Ignazio Pillai, Samuel Rota Bulò, Davide Ariu, Marcello Pelillo, and Fabio Roli. Is data clustering in adversarial settings secure? In Proceedings of the 2013 ACM workshop on Artificial intelligence and security , 2013. Nicholas Carlini. Poisoning the unlabeled dataset of semi-supervised learning. In 30th USENIX Security Symposium (USENIX Security 21) , 2021. Gal Chechik, Varun Sharma, Uri Shalit, and Samy Bengio. Large scale online learning of image similarity through ranking. Journal of Machine Learning Research , 11(36):1109–1135, 2010. URLhttp://jmlr.org/papers/v11/chechik10a.html . Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning , pp.

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

38

contrastive learning of visual representations. In International conference on machine learning , pp. 1597–1607. PMLR, 2020a. Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 , 2020b. Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 , 2017. Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05) , volume 1, pp. 539–546. IEEE, 2005. J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09 , 2009. Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

39

Image Database. In CVPR09 , 2009. Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. In Proceedings of the NIPS Workshop on Mach. Learn. and Comp. Sec , 2017. Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06) , volume 2, pp. 1735–1742. IEEE, 2006. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 770–778, 2016. Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 , 2019. Gabriel Ilharco, Mitchell Wortsman, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt.

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

40

Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. URL https://doi.org/10.5281/zenodo.5143773 . If you use this software, please cite it as below. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. arXiv preprint arXiv:2102.05918 , 2021. Armand Joulin, Laurens Van Der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In European Conference on Computer Vision , pp. 67–84. Springer, 2016. Marius Kloft and Pavel Laskov. Online anomaly detection under adversarial impact. In Proceedings of the thirteenth international conference on artificial intelligence and statistics , pp. 405–412, 2010. Marius Kloft and Pavel Laskov. Security analysis of online centroid anomaly detection. The Journal of Machine Learning Research , 13(1), 2012. 11

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

41

2010. Marius Kloft and Pavel Laskov. Security analysis of online centroid anomaly detection. The Journal of Machine Learning Research , 13(1), 2012. 11 Published as a conference paper at ICLR 2022 Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 , pp. 1885–1894. JMLR. org, 2017. Xuanqing Liu, Si Si, Xiaojin Zhu, Yang Li, and Cho-Jui Hsieh. A unified framework for data poisoning attack to graph-based semi-supervised learning. Advances in Neural Information Processing Systems , 2020. Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 , 2018. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

42

Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020 , 2021. Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning , pp. 5389–5400. PMLR, 2019. Aniruddha Saha, Ajinkya Tejankar, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Backdoor attacks on self-supervised learning, 2021. Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems , pp. 6103–6113, 2018. Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

43

hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 2556–2565, 2018. Richard Socher and Li Fei-Fei. Connecting modalities: Semi-supervised segmentation and annotation of images using unaligned text corpora. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition , pp. 966–973. IEEE, 2010. Kihyuk Sohn. Improved deep metric learning with multi-class n-pair loss objective. In Proceedings of the 30th International Conference on Neural Information Processing Systems , pp. 1857–1865, 2016. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations , 2014. URL http://arxiv.org/abs/1312.6199 . Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

44

Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems , 33, 2020. Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM , 59(2):64–73, 2016. Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849 , 2019. Yonglong Tian, Olivier J. Henaff, and Aaron van den Oord. Divide and contrast: Self-supervised learning from uncurated data. arXiv preprint arXiv:2105.08054 , 2021. Kerem Turgutlu. Self Supervised Learning with Fastai. Available from https:// keremturgutlu.github.io/self_supervised/ , 2021. Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771 , 2019. 12 Published as a conference paper at ICLR 2022

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

45

preprint arXiv:1912.02771 , 2019. 12 Published as a conference paper at ICLR 2022 Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762 , 2017. Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neighbor classification. Journal of machine learning research , 10(2), 2009. Jason Weston, Samy Bengio, and Nicolas Usunier. Large scale image annotation: learning to rank with joint word-image embeddings. Machine learning , 81(1):21–35, 2010. Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via nonparametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp. 3733–3742, 2018. Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. International Conference on Learning Representations , 2017.

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2106.09667

46

learning requires rethinking generalization. International Conference on Learning Representations , 2017. Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. arXiv preprint arXiv:2010.00747 , 2020. 13

2106.09667

Poisoning and Backdooring Contrastive Learning

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

http://arxiv.org/pdf/2106.09667

cs.LG

20210617

20220328

2005.12246

0

Demoting Racial Bias in Hate Speech Detection Mengzhou Xia Anjalie Field Yulia Tsvetkov Language Technologies Institute Carnegie Mellon University fmengzhox,anjalief,ytsvetko g@cs.cmu.edu Abstract In current hate speech datasets, there exists a high correlation between annotators’ perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification. 1 Introduction The prevalence of toxic comments on social media and the mental toll on human moderators has generated much interest in automated systems for detecting hate speech and abusive language (Schmidt and

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

1

1 Introduction The prevalence of toxic comments on social media and the mental toll on human moderators has generated much interest in automated systems for detecting hate speech and abusive language (Schmidt and Wiegand, 2017; Fortuna and Nunes, 2018), especially language that targets particular social groups (Silva et al., 2016; Mondal et al., 2017; Mathew et al., 2019). However, deploying these systems without careful consideration of social context can increase bias, marginalization, and exclusion (Bender and Friedman, 2018; Waseem and Hovy, 2016). Most datasets currently used to train hate speech classifiers were collected through crowdsourced annotations (Davidson et al., 2017; Founta et al., 2018), despite the risk of annotator bias. Waseem (2016) show that non-experts are more likely to label text as abusive than expert annotators, and Sap et al. (2019) show how lack of social context in annotation tasks further increases the riskof annotator bias, which can in turn lead to the marginalization of racial minorities. More specifically, annotators are more likely to label comments as abusive if they are written in African American English (AAE). These comments are assumed to be incorrectly labelled, as annotators do not mark

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

2

as abusive if they are written in African American English (AAE). These comments are assumed to be incorrectly labelled, as annotators do not mark them as abusive if they are properly primed with dialect and race information (Sap et al., 2019). These biases in annotations are absorbed and amplified by automated classifiers. Classifiers trained on biased annotations are more likely to incorrectly label AAE text as abusive than non-AAE text: the false positive rate (FPR) is higher for AAE text, which risks further suppressing an already marginalized community. More formally, the disparity in FPR between groups is a violation of the Equality of Opportunity criterion, a commonly used metric of algorithmic fairness whose violation indicates discrimination (Hardt et al., 2016). According to Sap et al. (2019), the false positive rate for hate speech/abusive language of the AAE dialect can reach as high as 46%. Thus, Sap et al. (2019) reveal two related issues in the task of hate speech classification: the first is biases in existing annotations, and the second is model tendencies to absorb and even amplify biases from spurious correlations present in datasets (Zhao et al., 2017; Lloyd, 2018). While current

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

3

is biases in existing annotations, and the second is model tendencies to absorb and even amplify biases from spurious correlations present in datasets (Zhao et al., 2017; Lloyd, 2018). While current datasets can be re-annotated, this process is timeconsuming and expensive. Furthermore, even with perfect annotations, current hate speech detection models may still learn and amplify spurious correlations between AAE and abusive language (Zhao et al., 2017; Lloyd, 2018). In this work, we present an adversarial approach to mitigating the risk of racial bias in hate speech classifiers, even when there might be annotation bias in the underlying training data. In x2, we describe our methodology in general terms, as it can be useful in any text classification task that seeksarXiv:2005.12246v1 [cs.CL] 25 May 2020 to predict a target attribute (here, toxicity) without basing predictions on a protected attribute (here, AAE). Although we aim at preserving the utility of classification models, our primary goal is not to improve the raw performance over predicting the target attribute (hate speech detection), but rather to reduce the influence of the protected attribute. Inx3 andx4, we evaluate how well our approach

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

4

target attribute (hate speech detection), but rather to reduce the influence of the protected attribute. Inx3 andx4, we evaluate how well our approach reduces the risk of racial bias in hate speech classification by measuring the FPR of AAE text, i.e., how often the model incorrectly labels AAE text as abusive. We evaluate our methodology using two types of data: (1) a dataset inferred to be AAE using demographic information (Blodgett et al., 2016), and (2) datasets annotated for hate speech (Davidson et al., 2017; Founta et al., 2018) where we automatically infer AAE dialect and then demote indicators of AAE in corresponding hate speech classifiers. Overall, our approach decreases the dialectal information encoded by the hate speech model, leading to a 2.2–3.2 percent reduction in FPR for AAE text, without sacrificing the utility of hate speech classification. 2 Methodology Our goal is to train a model that can predict a target attribute (abusive or not abusive language), but that does not base decisions off of confounds in data that result from protected attributes (e.g., AAE

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

5

that does not base decisions off of confounds in data that result from protected attributes (e.g., AAE dialect). In order to achieve this, we use an adversarial objective, which discourages the model from encoding information about the protected attribute. Adversarial training is widely known for successfully adapting models to learn representations that are invariant to undesired attributes, such as demographics and topics, though they rarely disentangle attributes completely (Li et al., 2018; Elazar and Goldberg, 2018; Kumar et al., 2019; Lample et al., 2019; Landeiro et al., 2019). Model Architecture Our demotion model consists of three parts: 1) An encoder Hthat encodes the text into a high dimensional space; 2) A binary classifierCthat predicts the target attribute from the input text; 3) An adversary Dthat predicts the protected attribute from the input text. We used a single-layer bidirectional LSTM encoder with an attention mechanism. Both classifiers are two-layer MLPs with a tanh activation function. Training Procedure Each data point in our training set is a tripletf(xi;yi;zi);i21:::Ng, wherexiis the input text, yiis the label for the target

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

6

attribute and ziis label of the protected attribute. The(xi;yi)tuples are used to train the classifier C, and the (xi;zi)tuple is used to train the adversary D. We adapt a two-phase training procedure from Kumar et al. (2019). We use this procedure because Kumar et al. (2019) show that their model is more effective than alternatives in a setting similar to ours, where the lexical indicators of the target and protected attributes are closely connected (e.g., words that are common in non-abusive AAE and are also common in abusive language datasets). In the first phase (pre-training), we use the standard supervised training objective to update encoder H and classifier C: min C;HNX i=1L(C(H(xi));yi) (1) After pre-training, the encoder should encode all relevant information that is useful for predicting the target attribute, including information predictive of the protected attribute. In the second phase, starting from the bestperforming checkpoint in the pre-training phase, we alternate training the adversary Dwith Equation 2 and the other two models ( HandC) with Equation 3: min D1 NNX

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

7

we alternate training the adversary Dwith Equation 2 and the other two models ( HandC) with Equation 3: min D1 NNX i=1L(D(H(xi));zi) (2) min H;C1 NNX i=1
L(C(H(xi));yi)+ (1)
L(D(H(xi));0:5)(3) Unlike Kumar et al. (2019), we introduce a hyper-parameter , which controls the balance between the two loss terms in Equation 3. We find thatis crucial for correctly training the model (we detail this inx3). We first train the adversary to predict the protected attribute from the text representations outputted by the encoder. We then train the encoder to “fool” the adversary by generating representations that will cause the adversary to output random guesses, rather than accurate predictions. At the same time, we train the classifier to predict the target attribute from the encoder output. Dataset Example Founta et al. (2018) I am hungry and I am dirty as hell bruh, need dat shower and dem calories

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

8

Dataset Example Founta et al. (2018) I am hungry and I am dirty as hell bruh, need dat shower and dem calories Blodgett et al. (2016) so much energy and time wasted hatin on someone when alla that coulda been put towards makin yourself better.... a https://t.co/awCg1nCt8t Table 1: Example from Founta et al. (2018) and Blodgett et al. (2016) where the state-of-the-art model misclassifies innocuous tweets (inferred to be AAE) as abusive language. Our model correctly classifies these tweets as nontoxic. 3 Experiments 3.1 Dataset To the best of our knowledge, there are no datasets that are annotated both for toxicity and for AAE dialect. Instead, we use two toxicity datasets and one English dialect dataset that are all from the same domain (Twitter): DWMW17 (Davidson et al., 2017) A Twitter dataset that contains 25K tweets annotated as hate speech ,offensive , ornone . The authors define hate speech as language that is used to expresses hatred towards a targeted group or is intended to be derogatory, to humiliate, or to insult the members

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

9

speech as language that is used to expresses hatred towards a targeted group or is intended to be derogatory, to humiliate, or to insult the members of the group, and offensive language as language that contains offensive terms which are not necessarily inappropriate. FDCL18 (Founta et al., 2018) A Twitter dataset that contains 100K tweets annotated as hateful ,abusive,spam ornone . This labeling scheme was determined by conducting multiple rounds of crowdsourcing to understand how crowdworkers use different labels. Strongly impolite, rude, or hurtful language is considered abusive, and the definition of hate speech is the same as in DWMW17. BROD16 (Blodgett et al., 2016) A 20K sample out of a 1.15M English tweet corpus that is demographically associated with African American twitter users. Further analysis shows that the dataset contains significant linguistic features of African American English. In order to obtain dialect labels for the DWMW17 and FDCL18, we use an off-the-shelf demographically-aligned ensemble model (Blodgett et al., 2016) which learns a posterior topic distribution (topics corresponding to African American, Hispanic, White and Other) at a user, message, and word level. Blodgett et al. (2016) generate a

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

10

distribution (topics corresponding to African American, Hispanic, White and Other) at a user, message, and word level. Blodgett et al. (2016) generate a AAE-aligned corpus comprising tweets from users labelled with at least 80% posterior probability asusing AAE-associated terms. Similarly, following Sap et al. (2019), we assign AAE label to tweets with at least 80% posterior probability of containing AAE-associated terms at the message level and consider all other tweets as Non-AAE. In order to obtain toxicity labels for the BROD16 dataset, we consider all tweets in this dataset to be non-toxic. This is a reasonable assumption since hate speech is relatively rare compared to the large amount of non-abusive language on social media (Founta et al., 2018).1 3.2 Training Parameters In the pre-training phase, we train the model until convergence and pick the best-performing checkpoint for fine-tuning. In the fine-tuning phase, we alternate training one single adversary and the classification model each for two epochs in one round and train for 10 rounds in total. We additionally tuned the parameter used to weight the loss terms in Equation 3 over validation

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

11

and train for 10 rounds in total. We additionally tuned the parameter used to weight the loss terms in Equation 3 over validation sets. We found that the value of is important for obtaining text representations containing less dialectal information. A large easily leads to over-fitting and a drastic drop in validation accuracy for hate speech classification. However, a near zeroseverely reduces both training and validation accuracy. We ultimately set = 0:05. We use the same architecture as Sap et al. (2019) as a baseline model, which does not contain an adversarial objective. For both of this baseline model and our model, because of the goal of demoting the influence of AAE markers, we select the model with the lowest false positive rate on validation set. We train models on both DWMW17 and FDCL18 datasets, which we split into train/dev/test subsets following Sap et al. (2019). 1We additionally did a simple check for abusive terms using a list of 20 hate speech words, randomly selected from Hatebase.org . We found that the percentage of sentences containing these words is much lower in AAE dataset ( 2%) than hate speech datasets ( 20%). Dataset Accuracy F1

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

12

containing these words is much lower in AAE dataset ( 2%) than hate speech datasets ( 20%). Dataset Accuracy F1 base ours base ours DWMW17 91.90 90.68 75.15 76.05 FDCL18 81.18 80.27 66.15 66.80 Table 2: Accuracy and F1 scores for detecting abusive language. F1 values are macro-averaged across all classification categories (e.g. hate, offensive, none for DWMW17). Our model achieves an accuracy and F1 on par with the baseline model. Offensive Hate base ours base ours FDCL18-AAE 20.94 17.69 3.23 2.60 BROD16 16.44 14.29 5.03 4.52 Table 3: False positive rates (FPR), indicating how often AAE text is incorrectly classified as hateful or abusive, when training with the FDCL18 dataset. Our model consistently improves FPR for offensiveness, and performs slightly better than the baseline for hate speech detection. 4 Results and Analysis Table 2 reports accuracy and F1 scores over the hate speech classification task. Despite the adversarial component in our model, which makes this task

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

13

4 Results and Analysis Table 2 reports accuracy and F1 scores over the hate speech classification task. Despite the adversarial component in our model, which makes this task more difficult, our model achieves comparable accuracy as the baseline and even improves F1 score. Furthermore, the results of our baseline model are on par with those reported in Sap et al. (2019), which verifies the validity of our implementation. Next, we assess how well our demotion model reduces the false positive rate in AAE text in two ways: (1) we use our trained hate speech detection model to classify text inferred as AAE in BROD16 dataset, in which we assume there is no hateful or offensive speech and (2) we use our trained hate speech detection model to classify the test partitions of the DWMW17 and FDCL18 datasets, which are annotated for hateful and offensive speech and for which we use an off-the-shelf model to infer dialect, as described in x3. Thus, for both evaluation criteria, we have or infer AAE labels and toxicity labels, and we can compute how often text inferred as AAE is misclassified as hateful, abusive, or offensive. Notably, Sap et al. (2019) show that datasets

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

14

often text inferred as AAE is misclassified as hateful, abusive, or offensive. Notably, Sap et al. (2019) show that datasets that annotate text for hate speech without sufficient context—like DWMW17 and FDCL18—may suffer from inaccurate annotations, in that annotatorsOffensive Hate base ours base ours DWMW17-AAE 38.27 42.59 0.70 2.06 BROD16 23.68 24.34 0.28 0.83 Table 4: False positive rates (FPR), indicating how often AAE text is incorrectly classified as hateful or offensive, when training with DWMW17 dataset. Our model fails to improve FPR over the baseline, since 97% of AAE-labeled instances in the dataset are also labeled as toxic. are more likely to label non-abusive AAE text as abusive. However, despite the risk of inaccurate annotations, we can still use these datasets to evaluate racial bias in toxicity detection because of our focus on FPR. In particular, to analyze false positives, we need to analyze the classifier’s predictions of the text as toxic, when annotators labeled it as non-toxic. Sap et al. (2019) suggest that annotators

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

15

of the text as toxic, when annotators labeled it as non-toxic. Sap et al. (2019) suggest that annotators over-estimate the toxicity in AAE text, meaning FPRs over the DWMW17 and FDCL18 test sets are actually lower-bounds, and the true FPR is could be even higher. Furthermore, if we assume that the DWMW17 and FDCL18 training sets contain biased annotations, as suggested by Sap et al. (2019), then a high FPR over the corresponding test sets suggests that the classification model amplifies bias in the training data, and labels non-toxic AAE text as toxic even when annotators did not. Table 3 reports results for both evaluation criteria when we train the model on the FDCL18 data. In both cases, our model successfully reduces FPR. For abusive language detection in the FDCL18 test set, the reduction in FPR is >3; for hate speech detection, the FPR of our model is also reduced by0:6compared to the baseline model. We can also observe a 2:2and0:5reduction in FPR for abusive speech and hate speech respectively when evaluating on BROD16 data. Table 4 reports results when we train the model

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

16

abusive speech and hate speech respectively when evaluating on BROD16 data. Table 4 reports results when we train the model on the DWMW17 dataset. Unlike Table 3, unfortunately, our model fails to reduce the FPR rate for both offensive and hate speech of DWMW17 data. We also notice that our model trained with DWMW17 performs much worse than the model trained with FDCL18 data. To understand the poor performance of our model when trained and evaluated on DWMW17 data, we investigated the data distribution in the test set and found that the vast majority of tweets 0 2 4 6 8 100.790.800.81 Accuracy 0 2 4 6 8 100.150.20 False Abusive Rate 0 2 4 6 8 10 Epoch0.020.04 False Hateful RateFigure 1: Accuracy of the entire development set of FDCL18 (top), and FPR rate for abusive (middle) and hate (bottom) speech detection for tweets inferred as AAE in the development set. X axis denotes the number of epochs. 0th epoch is the best checkpoint for pretraining step, which is also the baseline model. labeled as AAE by the dialect classifier were also annotated as toxic (97%). Thus, the subset of the data over which our model might improve FPR

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

17

labeled as AAE by the dialect classifier were also annotated as toxic (97%). Thus, the subset of the data over which our model might improve FPR consists of merely <3%of the AAE portion of the test set (49 tweets). In comparison, 70.98% of the tweets in the FDCL18 test set that were labeled as AAE were also annotated as toxic. Thus, we hypothesize that the performance of our model over the DWMW17 test set is not a representative estimate of how well our model reduces bias, because the improvable set in the DWMW17 is too small. In Table 1, we provide two examples of tweets that the baseline classifier misclassifies abusive/offensive, but our model, correctly classifies as non-toxic. Both examples are drawn from a toxicity dataset and are classified as AAE by the dialectal prediction model. Trade-off between FPR and Accuracy In order to better understand model performance, we explored the accuracy and FPR of our model throughout the entire training process. We evaluate the best checkpoint of the pre-trained model ( 0thepoch) and checkpoints of each epoch during adversarial training and show the results in Figure 1. While the

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

18

checkpoint of the pre-trained model ( 0thepoch) and checkpoints of each epoch during adversarial training and show the results in Figure 1. While the baseline model ( 0thepoch, before any adversarial training) achieves high accuracy, it also has a high FPR rate, particularly over abusive language. After adversarial training, the FPR rate decreases with only minor changes in accuracy. However, checkpoints with lower FPR rates also often have lower accuracy. While Tables 2 and 3 suggest that our model does achieve a balance between thesemetrics, Figure 1 shows the difficulty of this task; that is, it is difficult to disentangle these attributes completely. Eliminatation of protected attribute In Figure 2, we plot the validation accuracy of the adversary through the entire training process in order to verify that our model does learn a text representation at least partially free of dialectal information. Further, we compare using one adversary during training with using multiple adversaries (Kumar et al., 2019). Through the course of training, the validation accuracy of AAE prediction decreases by about 6–10 and 2–5 points for both datasets, indicating that dialectal information is gradually removed from the encoded representation. However, after a certain training threshold (6 epochs for

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

19

indicating that dialectal information is gradually removed from the encoded representation. However, after a certain training threshold (6 epochs for DWMW17 and 8 epochs for FDCL18), the accuracy of the classifier (not shown) also drops drastically, indicating that dialectal information cannot be completely eliminated from the text representation without also decreasing the accuracy of hatespeech classification. Multiple adversaries generally cause a greater decrease in AAE prediction than a single adversary, but do not necessarily lead to a lower FPR and a higher classification accuracy. We attribute this to the difference in experimental setups: in our settings, we focus on one attribute to demote, whereas Kumar et al. (2019) had to demote ten latent attributes and thus required multiple adversaries to stabilize the demotion model. Thus, unlike in (Kumar et al., 2019), our settings do not require multiple adversaries, and indeed, we do not see improvements from using multiple adversaries. 5 Related Work Preventing neural models from absorbing or even amplifying unwanted artifacts present in datasets is indispensable towards building machine learning systems without unwanted biases. One thread of work focuses on removing bias at the data level, through reducing annotator bias (Sap et al., 2019) and augmenting imbalanced datasets

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

20

One thread of work focuses on removing bias at the data level, through reducing annotator bias (Sap et al., 2019) and augmenting imbalanced datasets (Jurgens et al., 2017). Dixon et al. (2018) propose an unsupervised method based on balancing the training set and employing a proposed measurement for mitigating unintended bias in text classification models. Webster et al. (2018) present a gender-balanced dataset with ambiguous name-pair pronouns to provide diversity coverage for realworld data. In addition to annotator bias, sampling 0 5 10 15 200.700.750.80AAE Accuracy DWMW17 0 5 10 15 20 Epochs0.880.900.92AAE Accuracy FDCL18single adversary multiple adversariesFigure 2: Validation accuracy on AAE prediction of the adversary in the whole training process. The green line denotes the training setting of one adversary and the orange line denotes the training setting of multiple adversaries. strategies also result in topic and author bias in datasets of abusive language detection, leading to decreased classification performance when testing in more realistic settings, necessitating the adoption of cross-domain evaluation for fairness (Wiegand et al., 2019). A related thread of work on debiasing focuses

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

21

in more realistic settings, necessitating the adoption of cross-domain evaluation for fairness (Wiegand et al., 2019). A related thread of work on debiasing focuses at the model level (Zhao et al., 2019). Adversarial training has been used to remove protected features from word embeddings (Xie et al., 2017; Zhang et al., 2018) and intermediate representations for both texts (Elazar and Goldberg, 2018; Zhang et al., 2018) and images (Edwards and Storkey, 2015; Wang et al., 2018). Though previous works have documented that adversarial training fails to obliterate protected features, Kumar et al. (2019) show that using multiple adversaries more effectively forces the removal. Along similar lines, multitask learning has been adopted for learning task-invariant representations. Vaidya et al. (2019) show that multitask training on a related task e.g., identity prediction, allows the model to shift focus to toxic-related elements in hate speech detection. 6 Conclusion In this work, we use adversarial training to demote a protected attribute (AAE dialect) when training a classifier to predict a target attribute (toxicity). While we focus on AAE dialect and toxicity, our methodology readily generalizes to other settings,

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

22

a classifier to predict a target attribute (toxicity). While we focus on AAE dialect and toxicity, our methodology readily generalizes to other settings, such as reducing bias related to age, gender, orincome-level in any other text classification task. Overall, our approach has the potential to improve fairness and reduce bias in NLP models. 7 Acknowledgements We gratefully thank anonymous reviewers, Maarten Sap, and Dallas Card for their help with this work. The second author of this work is supported by the NSF Graduate Research Fellowship Program under Grant No. DGE1745016. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF. We also gratefully acknowledge Public Interest Technology University Network Grant No. NVF-PITU-Carnegie Mellon University-Subgrant-009246-2019-10-01 for supporting this research. References Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics , 6:587–604. Su Lin Blodgett, Lisa Green, and Brendan OConnor.

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

23

Transactions of the Association for Computational Linguistics , 6:587–604. Su Lin Blodgett, Lisa Green, and Brendan OConnor. 2016. Demographic dialectal variation in social media: A case study of african-american english. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages 1119–1130. Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Eleventh international aaai conference on web and social media . Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigating unintended bias in text classification. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society , pages 67–73. Harrison Edwards and Amos Storkey. 2015. Censoring representations with an adversary. arXiv preprint arXiv:1511.05897 . Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 11– 21.

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

24

removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 11– 21. Paula Fortuna and S ´ergio Nunes. 2018. A survey on automatic detection of hate speech in text. ACM Computing Surveys (CSUR) , 51(4):85. Antigoni Maria Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of twitter abusive behavior. In Twelfth International AAAI Conference on Web and Social Media . Moritz Hardt, Eric Price, and Nathan Srebro. 2016. Equality of opportunity in supervised learning. In Proceedings of the 30th International Conference on Neural Information Processing Systems , pages 3323–3331. David Jurgens, Yulia Tsvetkov, and Dan Jurafsky. 2017. Incorporating dialectal variability for socially equitable language identification. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , volume 2, pages 51–57.

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

25

55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , volume 2, pages 51–57. Sachin Kumar, Shuly Wintner, Noah A Smith, and Yulia Tsvetkov. 2019. Topics to avoid: Demoting latent confounds in text classification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages 4144–4154. Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc’Aurelio Ranzato, and YLan Boureau. 2019. Multiple-attribute text rewriting. In International Conference on Learning Representations . Virgile Landeiro, Tuan Tran, and Aron Culotta. 2019. Discovering and controlling for latent confounds in text classification using adversarial domain adaptation. In Proceedings of the 2019 SIAM International Conference on Data Mining , pages 298–305. SIAM. Yitong Li, Timothy Baldwin, and Trevor Cohn. 2018. Towards robust and privacy-preserving text representations. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

26

Towards robust and privacy-preserving text representations. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages 25–30, Melbourne, Australia. Association for Computational Linguistics. Kirsten Lloyd. 2018. Bias amplification in artificial intelligence systems. CoRR , abs/1809.07842. Binny Mathew, Ritam Dutt, Pawan Goyal, and Animesh Mukherjee. 2019. Spread of hate speech in online social media. In Proceedings of the 10th ACM Conference on Web Science , pages 173–182. ACM. Mainack Mondal, Leandro Ara ´ujo Silva, and Fabr ´ıcio Benevenuto. 2017. A measurement study of hate speech in social media. In Proceedings of the 28th ACM Conference on Hypertext and Social Media , pages 85–94. ACM. Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the57th Annual Meeting of the Association for Computational Linguistics , pages 1668–1678. Anna Schmidt and Michael Wiegand. 2017. A survey

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

27

Anna Schmidt and Michael Wiegand. 2017. A survey on hate speech detection using natural language processing. In Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media , pages 1–10, Valencia, Spain. Association for Computational Linguistics. Leandro Silva, Mainack Mondal, Denzil Correa, Fabr´ıcio Benevenuto, and Ingmar Weber. 2016. Analyzing the targets of hate in online social media. In Tenth International AAAI Conference on Web and Social Media . Ameya Vaidya, Feng Mai, and Yue Ning. 2019. Empirical analysis of multi-task learning for reducing model bias in toxic comment detection. arXiv preprint arXiv:1909.09758 . Tianlu Wang, Jieyu Zhao, Kai-Wei Chang, Mark Yatskar, and Vicente Ordonez. 2018. Adversarial removal of gender from deep image representations. arXiv preprint arXiv:1811.08489 . Zeerak Waseem. 2016. Are you a racist or am i seeing things? annotator influence on hate speech detection on twitter. In Proceedings of the first workshop on NLP and computational social science , pages 138– 142.

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

28

on twitter. In Proceedings of the first workshop on NLP and computational social science , pages 138– 142. Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on twitter. In Proceedings of the NAACL student research workshop , pages 88–93. Kellie Webster, Marta Recasens, Vera Axelrod, and Jason Baldridge. 2018. Mind the gap: A balanced corpus of gendered ambiguous pronouns. Transactions of the Association for Computational Linguistics, 6:605–617. Michael Wiegand, Josef Ruppenhofer, and Thomas Kleinbauer. 2019. Detection of abusive language: the problem of biased datasets. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 602–608. Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, and Graham Neubig. 2017. Controllable invariance through adversarial feature learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems , pages 585–596. Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. 2018. Mitigating unwanted biases with

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2005.12246

29

of the 31st International Conference on Neural Information Processing Systems , pages 585–596. Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. 2018. Mitigating unwanted biases with adversarial learning. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society , pages 335–340. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Ryan Cotterell, Vicente Ordonez, and Kai-Wei Chang. 2019. Gender bias in contextualized word embeddings. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 629–634. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing , pages 2979–2989.

2005.12246

Demoting Racial Bias in Hate Speech Detection

In current hate speech datasets, there exists a high correlation between annotators' perceptions of toxicity and signals of African American English (AAE). This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by current hate speech classifiers. In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. Experimental results on a hate speech dataset and an AAE dataset suggest that our method is able to substantially reduce the false positive rate for AAE text while only minimally affecting the performance of hate speech classification.

http://arxiv.org/pdf/2005.12246

Accepted at SocialNLP Workshop @ACL 2020

cs.CL

20200525

20200525

2008.02637

0

arXiv:2008.02637v1 [cs.CL] 6 Aug 2020Question and Answer Test-Train Overlap in Open-Domain Ques tion Answering Datasets Patrick Lewis†‡, Pontus Stenetorp‡, Sebastian Riedel†‡ †Facebook AI Research;‡University College London plewis@fb.com Abstract Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

1

test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models outperform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks. 1 Introduction Open-domain Question Answering (ODQA) is a task examining the ability of models to produce answers to natural language factoid questions drawn from an open set of domains. ODQA has received significant attention for its potential practical applications, and more recently as a popular method to analyse how well NLP systems can capture and recall factual knowledge. This interest in ODQA as a challenging “knowledge-intensive” task has led to a flurry of recent works that have driventest-set performance on standard ODQA datasets

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

2

as a challenging “knowledge-intensive” task has led to a flurry of recent works that have driventest-set performance on standard ODQA datasets to new heights ( Lee et al. ,2019 ;Guu et al. ,2020 ; Karpukhin et al. ,2020 ;Lewis et al. ,2020 ;Izacard and Grave ,2020 , inter alia). However, a deeper understanding of what kinds of questions our models can answer well has been less forthcoming. Whilst there have been several works examining other kinds of QA datasets ( Manjunatha et al. ,2018 ; Kaushik and Lipton ,2018 ;Sugawara et al. ,2018 , 2020 ), we know comparatively little about how the questions and answers are distributed in these ODQA benchmarks, making it hard to understand and contextualize the results we are observing. In this work, we address these issues via an analysis of the test sets of three popular ODQA datasets, namely WebQuestions ( Berant et al. , 2013 ), TriviaQA ( Joshi et al. ,2017 ) and Open Natural Questions ( Kwiatkowski et al. ,2019 ;Lee et al. ,2019 ). We identify three classes of question

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

3

Natural Questions ( Kwiatkowski et al. ,2019 ;Lee et al. ,2019 ). We identify three classes of question that a trained ODQA system should be able to answer, in increasing order of difficulty: 1) the most basic behaviour is to be able to reliably recall the answer to a question that the model has seen at training time. 2) a model should be able to answer novel questions at test time and choose an answer from the set of answers it has seen during training. 3) a strong system should be able to answer novel questions which have answers which are not contained in the training data. It is not clear to what extent our current ODQA datasets measure each of these three behaviours. To address this, we stratify the test sets of these datasets. Firstly, we split the test data by whether answers in the test set also appear somewhere in the training sets. We find that 58-71% of test answers also occur somewhere in the training data, demonstrating that the majority of the test data does not probe for answer generalization. Secondly, we annotate 1000 question, answer pairs from each test set for repeated questions in Dataset% Answer overlap% Question overlap Natural Questions 63.6 32.5 TriviaQA 71.7 33.6

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

4

Dataset% Answer overlap% Question overlap Natural Questions 63.6 32.5 TriviaQA 71.7 33.6 WebQuestions 57.9 27.5 Table 1: Fractions of open-domain test sets that overlap with their training sets. their respective training sets. We find that a surprisingly high 28-34% have paraphrased questions in the training data, the vast majority of which are near-duplicates differing by one or two words. This result implies that 30% of the test set of these datasets only probe for how well models can simply memorize question answer pairs seen at training. Equipped with these insights, we compute the performance of several recently proposed ODQA models on our test subsets. We test both Open-book approaches, which leverage retrieval from a large corpus of documents and Closedbook approaches, which focus on training large parametric models with no external knowledge source ( Roberts et al. ,2020 ). We find that test data with train-overlapping data contribute the bulk of the overall performance of all the models studied. These issues seem to be more acute for closedbook models. Strikingly, we find that a closedbook BART-based model ( Lewis et al. ,2019 ) is incapable of producing answers not observed at

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

5

incapable of producing answers not observed at training time, and achieves very low scores on non-overlapping questions, suggesting this model is only capable of memorizing question, answer pairs from training time. With this in mind, we build simple nearest-neighbor models which outperform this BART model, despite having virtually no capacity to generalize beyond training data. To summarize, we make the following contributions: 1) We provide insights into how answer entities are distributed between dataset splits for ODQA datasets 2) We provide annotated subsets of ODQA test sets indicating whether test-time questions are duplicates of training time questions.13) We evaluate a variety of models on our dataset splits, and derive insights into what kinds of question answering behaviour different models achieve. 1Our data and evaluation code will be made available at https://github.com/facebookresearch/QA-Overlap2 Datasets In our analysis, we consider three widely used Open-domain QA datasets, WebQuestions ( Berant et al. ,2013 ), TriviaQA ( Joshi et al. ,2017 ), and Open Natural Questions, a subset of Natural Questions ( Kwiatkowski et al. ,2019 ) introduced by Lee et al. (2019 ). All three datasets consist of factual natural language questions and short multi-token

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

6

et al. (2019 ). All three datasets consist of factual natural language questions and short multi-token answers, but differ slightly in the style of questions and format of answers. WebQuestions WebQuestions is a dataset of 3,778 train and 2,032 test question, answer pairs. Questions were obtained by mining a search engine, and answers are Freebase entities ( Bollacker et al. ,2008 ) annotated by crowdworkers. The ODQA task consists of predicting the name of the freebase entity. We use the standard train/test splits from Berant et al. (2013 ). We use the development split used in Karpukhin et al. (2020 ), which was randomly split from the train set. TriviaQA TriviaQA is a dataset of 78,785 train, 8,837 development and 11,313 test question, answer pairs obtained by scraping trivia websites. Answers consist of wikipedia entities, and any alias for the answer entity is considered a correct answer. We use the open-domain train/test splits, which corresponding to the unfiltered-train and unfiltered-dev reading comprehension splits ( Lee et al. ,2019 ;Min et al. ,2019 ,2020 ;Karpukhin et al. ,2020 ). Open-Natural Questions Natural Questions

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

7

et al. ,2019 ;Min et al. ,2019 ,2020 ;Karpukhin et al. ,2020 ). Open-Natural Questions Natural Questions consists of search engine questions with answers annotated as spans in wikipedia articles by crowdworkers. The open-domain version of the dataset consists of question, answer pairs from Natural Questions which have short answer spans less than 6 tokens in length. We use the standard opendomain splits in our experiments, consisting of 79,168 train, 8,757 development and 3,610 question answer pairs. For all three datasets, the canonical train, development and test splits were obtained by randomly splitting the question, answer pairs, and there are no exact duplicate questions in any dataset. We exclude development data from our overlap analyses, and focus purely on train-test overlap to explicitly assess the effects of training memorization. 3 Test-Train Overlaps We explore two ways of examining the test sets based on overlaps between training and test data. Consider a question, answer pair (q,a)from the test setDtestwhere the answer consists of at least one answer reference a={s1..sn}. We can consider answer overlap where there exists at least one(q′,a′)∈Dtrainwhich shares at least one

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

8

one answer reference a={s1..sn}. We can consider answer overlap where there exists at least one(q′,a′)∈Dtrainwhich shares at least one answer reference with (q,a). We can also consider question overlap , where there exists some (q′′,a′′)∈Dtrainwhereq′′is a duplicate of q, such that qandq′′are paraphrases and have the same answer. Answer Overlap Following Rajpurkar et al. (2016 ), we apply answer normalization2on answer references before searching for overlapping answer references for all (q,a)pairs in the test set – see Table 1. We find that 58% of test (q,a) pairs in WebQuestions have answer overlaps, with 63.6% and 71.7% for Natural Questions and TriviaQA respectively. We would naturally expect TriviaQA to have higher answer overlap as it has more answer references per question on average (13.7 references on average compared to 1.2 for Natural Questions and 2.4 for WebQuestions). Examples of answer overlaps are shown in Table 2. Question-Overlap Unlike answer overlap, question overlap cannot be easily computed automatically, as searching for duplicates via rules or paraphrase classifiers may lead to both false positives

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

9

Question-Overlap Unlike answer overlap, question overlap cannot be easily computed automatically, as searching for duplicates via rules or paraphrase classifiers may lead to both false positives and negatives. Thus, we turn to manual annotation to investigate question overlap. To obtain a representative sample for each dataset, we annotate a random subset of 1,000 (q,a)pairs for each test set. Annotators are shown a list of up to 50 training questions which have a similar answer reference.3This answer similarity function is designed for high recall to obtain a tight lower bound on question overlap. If there were no questions with similar answers in the training set, the question was automatically annotated as not over2Answer normalization consists of lower-casing, stripping punctuation, removing articles and normalizing white space 3Training questions are selected for annotation if one of the following is true: they share an answer reference with a test question, a test answer reference is a sub-sequence of a training answer reference, or the other way around (a training reference answer is a sub-sequence of a test answer reference). If there are more than 50 such questions, the top 50 are chosen by the highest degree of word overlap to the test question.lapping. Three expert annotators looked through these similar questions and indicated if any were paraphrases of the test question and had the same

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

10

question.lapping. Three expert annotators looked through these similar questions and indicated if any were paraphrases of the test question and had the same answer. The results from the annotation can be seen in Table 1shows these results and examples of overlapping questions in Table 3. A sample of 100 2-way annotated examples indicated 93% agreement, corresponding to a Cohen’s Kappa of 0.85 ( Cohen ,1960 ). What we observe is a high degree of question overlap, with between 27.5 and 33.6% of the 1,000 annotated test questions had a duplicate in the training set. It is also common to see several duplicates per test question, with an average of 2.8 duplicate questions per overlapping test question in Natural Questions. 4 Implications for Modelling Given our findings from above, we turn our attention to how well ODQA models perform with respect to train-test set overlap. Earlier, we identified three classes of answering behaviors: 1) questions that can be memorized at training time, 2) novel questions that can be answered with answers memorized at training time, 3) novel questions with novel answers. We refer to these behaviours asQuestion memorization ,Answer classification andQA generalization respectively.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

11

with novel answers. We refer to these behaviours asQuestion memorization ,Answer classification andQA generalization respectively. Question memorization To perform well on the question overlap subset, a model would only need to be able to memorize (q,a)pairs at training time, then recognize which training question matches a test-time question. The reasoning required ranges from trivial duplicate detection for very similar questions such as “who played pink in pink floyd the wall” and “who played pink in the movie the wall”, to more challenging inference problems for more subtle duplicates such as “On which island in the North Sea did both St Aidan and St Cuthbert live?” and “irish born missionary saint aidan founded a monastery in 653 on which english island which is also the name of a 1970s uk folk-rock band?”. A manual annotation of 100 question-overlap pairs indicated that 81% were simple duplicates differing by one or two words, 14% required some paraphrasing recognition capability, and 5% required more sophisticated natural language understanding. To measure performance on question memorization, we build a test subset comprised of (q,a)pairs which have question overlap to the training set. Open Natural Questions TriviaQA WebQuestions

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

12

subset comprised of (q,a)pairs which have question overlap to the training set. Open Natural Questions TriviaQA WebQuestions Overlapping Non-overlapping Overlapping Non-overlapping Overlapping Non-overlapping Phil Simms Cloves David Bowie Death in the afternoon Harvard Queen Victoria Brian Johnson Matt Monro Battle of camlann Clash of the Titans Alderaan Braslia 8 1,020 – 1,080 kg Heligoland ice-cream sundae India Paddington the Indians Hermann Ebbinghaus Henry VII Camshaft 2011 Tom Corbett the 1830s Matt Flinders Niagra Falls Cumberland Zeus Gary Table 2: Randomly sampled overlapping and non-overlapping answers from all three test sets. Answer Test Question Train Question Jason Marsden who plays max voice in a goofy movie who does max voice in a goofy movie January 23 2018 when will the 2018 oscar nominations be annou nced when are the oscar nominations for 2018 announced Alan Shearer who has scored more goals in the premier league m ost goals scored by a premier league player retina where are the cones in the eye located where are cone ce lls located in the eye francisco pizarro who led the conquest of the incas in south a merica conquistador who defeated the incan empire in peru

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

13

francisco pizarro who led the conquest of the incas in south a merica conquistador who defeated the incan empire in peru Table 3: Randomly sampled test-train overlapping question s in Open Natural Questions. See Appendix A.1for more examples, including examples from TriviaQA and WebQue stions Answer Classification In order to tackle the answer-overlap question, a multi-class classifier over training set answers would be sufficient, as answers never appear at test time that don’t appear at training time. We build a test subset of (q,a) pairs which have answer overlap, but do not have question overlap. Question-overlap pairs are excluded to isolate performance on answer classification, since question-overlap questions are significantly easier to answer, and would inflate scores. QA Generalization In this regime, models cannot rely on memorizing their training data. To measure performance on this most challenging split, we build a test subset of (q,a)pairs which do not have answer overlap with the training set. We further note that we expect higher frequency

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

14

split, we build a test subset of (q,a)pairs which do not have answer overlap with the training set. We further note that we expect higher frequency answers, such as countries, integers and public figures would naturally be expected to appear less often in this test subset. As such, models that perform well on the head of the answer distribution may struggle to perform well in this setting, despite being able to perform some generalization at test time. In the following, we briefly describe the models included in our analysis. For published models, we obtain test set predictions directly from the authors. 4.1 Open-Book Models Open-book Models are ODQA models which first retrieve relevant documents from Wikipedia and then either extract or generate answers conditioned on those documents. We consider theDense Passage Retrieval (DPR) model ( Karpukhin et al. ,2020 ), a pipeline model which retrieves documents based on dense embeddings, before feeding them into a conventional reader-reranker which extracts spans of text as answers. We also include Retrieval-Augmented Generation ( Lewis et al. ,2020 ), a recent model that jointly learns to retrieve and generate answers in seq2seq framework, based on dense retrieval and BART ( Lewis

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

15

et al. ,2020 ), a recent model that jointly learns to retrieve and generate answers in seq2seq framework, based on dense retrieval and BART ( Lewis et al. ,2019 ). Finally we include the state-of-theart Fusion-in-Decoder (FID) ( Izacard and Grave , 2020 ), a pipeline model based on T5-large ( Raffel et al. ,2020 ) which retrieves 100 documents and fuses them so that the decoder can attend to all documents at once. We not include FID results on WebQuestions as the authors did not use it in their original work. 4.2 Closed-Book Models Closed-book models store the knowledge required to answer their questions entirely within the parameters of the model itself, rather than in an external corpus. Typically these models consist of seq2seq transformer models which are directly fine-tuned on (q,a)pairs. In our analysis, we train a BART-large closed-book QA model, which is trained with questions as input and generates (q,a)pairs as output. Checkpoints are selected by Exact Match score on a development set. We also include a much more powerful T5-11B model

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

16

(q,a)pairs as output. Checkpoints are selected by Exact Match score on a development set. We also include a much more powerful T5-11B model from Roberts et al. (2020 ). We use the T511B model which has been pretrained with a special “Salient Span Masking” objective ( Guu et al. , 2020 ), designed to improve downstream ODQA ModelOpen Natural Questions TriviaQA WebQuestions TotalQuestion OverlapAnswer Overlap OnlyNo OverlapTotalQuestion OverlapAnswer Overlap OnlyNo OverlapTotalQuestion OverlapAnswer Overlap OnlyNo Overlap Open bookRAG 44.5 70.7 34.9 24.8 56.8 82.7 54.7 29.2 45.5 81.0 45.8 21.1 DPR 41.3 69.4 34.6 19.3 57.9 80.4 59.6 31.6 42.4 74.1 39.8 22.2 FID 51.4 71.3 48.3 34.5 67.6 87.5 66.9 42.8 - - - Closed bookT5-11B+SSM 36.6 77.2 22.2 9.4 - - - - 44.7 82.1 44.5 22.0

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

17

bookT5-11B+SSM 36.6 77.2 22.2 9.4 - - - - 44.7 82.1 44.5 22.0 BART 26.5 67.6 10.2 0.8 26.7 67.3 16.3 0.8 27.4 71.5 20.7 1.6 Nearest NeighborDense 26.7 69.4 7.0 0.0 28.9 81.5 11.2 0.0 26.4 78.8 17.1 0.0 TF-IDF 22.2 56.8 4.1 0.0 23.5 68.8 5.1 0.0 19.4 63.9 8.7 0.0 Table 4: Exact Match scores for several recent models on our d ataset splits. The “Total” column is the overall performance on the dataset. “Question Overlap” refers to th e test subset with train-test question overlap, and probes for simple question memorization. “Answer Overlap O nly” refers to the test subset without train-test question overlap, but with train-test answer overlap, whic h probes for answer classification. “No overlap” refers to the test subset with no train-test answer overlap and probes for QA generalization

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

18

the test subset with no train-test answer overlap and probes for QA generalization performance. The T5-11B model was trained on both train and development portions of the data, and thus has seen ∼10% more training data than other models. As we did not include development data in our overlap analysis, a small amount of unaccounted-for overlap is possible for this model. We do not include TriviaQA results for the T5 model since this model was trained using a different TriviaQA data splitting scheme. 4.3 Nearest-Neighbor Models Given that there are high levels of train-test overlaps in these datasets, we also experiment with some simple nearest-neighbor models. Here, we simply retrieve a (q,a)pair from the training set based on question similarity to the test question, and return its answer. We experiment with two models, one using TF-IDF and the other using the dot product similarity of question embeddings from the DPR retriever. These models cannot generalize to non-overlapping answers, and have limited capacity to answer non-overlapping questions. However, these models are attractive from the perspective of model size and efficiency. There has recently been a push towards more space and

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

19

the perspective of model size and efficiency. There has recently been a push towards more space and memory-efficient QA systems.4Lightweight retrievers coupled with a database of carefully selected(q,a)pairs would represent a very spaceefficient solution compared to open-book models which must retrieve from large textual corpora, or closed-book models with large parameter counts. 4Such as the EfficientQA competition at Neurips 2020 https://efficientqa.github.io/4.4 Results Table 4shows our results. In this section we unpack some findings. Question Memorization Earlier, we found that ∼30% of test set questions overlap with the training set. The “Question overlap” columns in Table4shows performance on Question Memorization. Comparing this column with the total performance column shows that all models perform significantly higher on memorizable questions. This finding is not surprising, but it is worth highlighting that a significant proportion of overall performance is driven by question memorization. This effect is most pronounced for closed book models. The T5-11B performs especially well for question

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

20

effect is most pronounced for closed book models. The T5-11B performs especially well for question memorization on both Natural Questions and WebQuestions. This suggests that its very large capacity, coupled with more powerful question understanding may allow it to store, recognise and recall training questions more effectively than other models. Answer Classification The “Answer overlap only” column in Table 4shows performance on answer classification. Answer classification has a large drop in performance compared to question memorization, dropping by an average of 45% Exact Match score. Open-book models handle this setting better than closed book models. The BART model in particular struggles here, only managing 10.2% accuracy on this set. QA Generalization The “No overlap” column in Table 4shows performance on QA generalization. All models suffer significant performance degradation on QA generalization, highlighting the shortcomings of the overall performance metric. For example, we may expect the FID stateof-the model to answer half of Natural Questionsstyle questions correctly, but once we have accounted for repeated questions and answers, it can only answer about one third of questions correctly. This difference is even more pronounced for other models, with an average absolute drop of 25% with respect to overall performance.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

21

only answer about one third of questions correctly. This difference is even more pronounced for other models, with an average absolute drop of 25% with respect to overall performance. Nearest-Neighbor Models The bottom two rows of Table 4show the results of our nearestneighbor models. The TF-IDF model, despite being completely untrained, is able to answer about 20% of test questions correctly, purely by retrieving questions from the training sets. More interestingly, the dense retrieval model outperforms the BART open-domain QA model on Natural Questions and TriviaQA. Furthermore, the dense nearest neighbor model also outperforms the significantly more complex DPR open-book model on TriviaQA and WebQuestions on the question overlap subset. These models have limitations, but represent very space and memory efficient solutions. Our dense nearest neighbour model consists of a single BERT-base checkpoint and outperforms a BART-large model, and could be compressed using quantization and distillation techniques ( Sanh et al. ,2020 ;Jiao et al. ,2019 ). The TF-IDF model is even smaller and could be implemented extremely efficiently with negligible memory footprint. 5 Related Work The widespread adoption of deep learning in the last few years has been accompanied with an increase in dataset sizes and construction methodologies. Examining what kinds of behaviours are

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

22

5 Related Work The widespread adoption of deep learning in the last few years has been accompanied with an increase in dataset sizes and construction methodologies. Examining what kinds of behaviours are learnt by models has received attention in natural langauge understanding tasks, such as the GLUE benchmark ( Wang et al. ,2018 ), which includes a diagnostic test set probing for different reasoning types. Various works have also performed critical and careful analysis of question answering systems and datasets. Chen et al. (2016 ) closely examine the difficulty of the CNN-DM dataset ( Hermann et al. ,2015 ),Sugawara et al. (2020 ) perform an analysis of machine comprehension dataset difficulty, Kaushik and Lipton (2018 ) analyse the difficulty of various machine reading datasets, and Manjunatha et al. (2018 ) show that visual question answering models memorize common question-answer relationships present in training data. Fvry et al. (2020 ) perform an analysis of various closedbook models’ TriviaQA predictions, based on entity mentions. Kwiatkowski et al. (2019 ) note that the machine reading Natural Questions dataset has substantial train-test overlap of wikipedia titles, and provide some baselines for “long-answer”

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

23

the machine reading Natural Questions dataset has substantial train-test overlap of wikipedia titles, and provide some baselines for “long-answer” QA. Closest to our work, Verga et al. (2020 ) observe similar answer overlap in knowledge-base QA, and explore results on non-overlapping subsets. 6 Conclusion In this work, we performed a novel analysis of popular open-domain question answering datasets. We found that 60% of test set answers overlap with the training set and, more surprisingly, 30% of test set questions have at least one duplicate in the train set. Following these observations, we contextualize the performance of seven ODQA models, stratifying by different amounts of training set overlap, gaining an insight into to what extent these models generalize or simply memorize their training data. It is clear that performance on these datasets cannot be properly understood by overall QA accuracy and suggest that in future, a greater emphasis should be placed on more behaviour-driven evaluation, rather than pursuing single-number overall accuracy figures. Acknowledgments The authors would like to thank thank Nicola De Cao, Tom Kwiatkowski, Michael Collins, Kenton Lee, Adam Roberts, Colin Raffel, Scott Yih, Sewon Min, Gautier Izacard and Vladimir

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

24

Kenton Lee, Adam Roberts, Colin Raffel, Scott Yih, Sewon Min, Gautier Izacard and Vladimir Karpuhkin for helpful discussions and providing test set prediction files for analysis. References Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic Parsing on Freebase from Question-Answer Pairs . InProceedings of the 2013 Conference on Empirical Methods in Natural Language Processing , pages 1533–1544, Seattle, Washington, USA. Association for Computational Linguistics. Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge . In Proceedings of the 2008 ACM SIGMOD international conference on Management of data , SIGMOD ’08, pages 1247–1250, Vancouver, Canada. Association for Computing Machinery. Danqi Chen, Jason Bolton, and Christopher D. Manning. 2016. A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task . In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 2358–2367, Berlin, Germany. Association for Computational Linguistics.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

25

Long Papers) , pages 2358–2367, Berlin, Germany. Association for Computational Linguistics. Jacob Cohen. 1960. A Coefficient of Agreement for Nominal Scales .Educational and Psychological Measurement , 20(1):37–46. Publisher: SAGE Publications Inc. Thibault Fvry, Livio Baldini Soares, Nicholas FitzGerald, Eunsol Choi, and Tom Kwiatkowski. 2020. Entities as Experts: Sparse Memory Access with Entity Supervision .arXiv:2004.07202 [cs] . ArXiv: 2004.07202. Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. REALM: Retrieval-Augmented Language Model Pre-Training . arXiv:2002.08909 [cs] . ArXiv: 2002.08909. Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching Machines to Read and Comprehend . In C. Cortes, N. D.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

26

to Read and Comprehend . In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28 , pages 1693–1701. Curran Associates, Inc. Gautier Izacard and Edouard Grave. 2020. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering .arXiv:2007.01282 [cs]. ArXiv: 2007.01282. Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2019. TinyBERT: Distilling BERT for Natural Language Understanding .arXiv:1909.10351 [cs] . ArXiv: 1909.10351. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension . InProceedings of the 55th AnnualMeeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 1601–1611, Vancouver, Canada. Association for Computational Linguistics.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

27

Vancouver, Canada. Association for Computational Linguistics. Vladimir Karpukhin, Barlas Ouz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering . arXiv:2004.04906 [cs] . ArXiv: 2004.04906. Divyansh Kaushik and Zachary C. Lipton. 2018. How Much Reading Does Reading Comprehension Require? A Critical Investigation of Popular Benchmarks . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 5010–5015, Brussels, Belgium. Association for Computational Linguistics. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: a Benchmark for Question Answering

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

28

Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: a Benchmark for Question Answering Research .Transactions of the Association of Computational Linguistics . Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent Retrieval for Weakly Supervised Open Domain Question Answering . InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 6086–6096, Florence, Italy. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension .arXiv:1910.13461 [cs, stat] . ArXiv: 1910.13461. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kttler, Mike Lewis, Wen-tau Yih, Tim Rocktschel, Sebastian Riedel, and Douwe Kiela. 2020.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

29

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks .arXiv:2005.11401 [cs] . ArXiv: 2005.11401. Varun Manjunatha, Nirat Saini, and Larry S. Davis. 2018. Explicit Bias Discovery in Visual Question Answering Models .arXiv:1811.07789 [cs] . ArXiv: 1811.07789. Sewon Min, Danqi Chen, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2019. A Discrete Hard EM Approach for Weakly Supervised Question Answering . InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages 2851– 2864, Hong Kong, China. Association for Computational Linguistics. Sewon Min, Danqi Chen, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2020. Knowledge Guided Text Retrieval and Reading for Open Domain Question Answering .arXiv:1911.03868 [cs] . ArXiv: 1911.03868.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

30

Retrieval and Reading for Open Domain Question Answering .arXiv:1911.03868 [cs] . ArXiv: 1911.03868. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text– to-Text Transformer .Journal of Machine Learning Research , 21(140):1–67. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages 2383–2392, Austin, Texas. Association for Computational Linguistics. Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How Much Knowledge Can You Pack Into the Parameters of a Language Model? arXiv:2002.08910 [cs, stat] . ArXiv: 2002.08910. Victor Sanh, Lysandre Debut, Julien Chaumond, and

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

31

[cs, stat] . ArXiv: 2002.08910. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2020. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter . arXiv:1910.01108 [cs] . ArXiv: 1910.01108. Saku Sugawara, Kentaro Inui, Satoshi Sekine, and Akiko Aizawa. 2018. What Makes Reading Comprehension Questions Easier? InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 4208–4219, Brussels, Belgium. Association for Computational Linguistics. Saku Sugawara, Pontus Stenetorp, Kentaro Inui, and Akiko Aizawa. 2020. Assessing the Benchmarking Capacity of Machine Reading Comprehension Datasets . In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The ThirtySecond Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020 , pages 8918–8927. AAAI Press.

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

32

7-12, 2020 , pages 8918–8927. AAAI Press. Pat Verga, Haitian Sun, Livio Baldini Soares, and William W. Cohen. 2020. Facts as Experts: Adaptable and Interpretable Neural Memory over Symbolic Knowledge .arXiv:2007.00849 [cs] . ArXiv: 2007.00849. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding . In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP , pages 353–355, Brussels, Belgium. Association for Computational Linguistics.A Appendices A.1 Additional Question Overlap Examples Tables 5,6and7give more question overlap examples for the three datasets. Answer Test Question Train Question Bob Geldof who played pink in pink floyd the wall who played pin k in the movie the wall Daren Maxwell Kagasoffwho played ricky in secret life of the american teenager who p layed ricky on the secret life of the american teenager Andy who does april end up with on parks and rec who does april m arry in parks and rec

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

33

Andy who does april end up with on parks and rec who does april m arry in parks and rec may 5 2017 when did gaurdians of the galaxy 2 come out when is gu ardians of the galaxy vol 2 released norman pritchard who won the first medal in olympics for india who won the first individual olympic medal for india moira kelly who does the voice of nala in the lion king who play ed nala in the lion king movie supreme court who enforces the charter of rights and freedom s who has final authority of interpretation of the canadian charter of rights and freedoms 554 most passing yards by nfl qb in a game what is the nfl record fo r most passing yards in a single game John Ross who ran the fastest 40 yard dash in the nfl who has the f astest 40 yard dash ever international border ib what is the name of india pakistan bo rder what is the border name between india and pakistan Andrew Wright who wrote when a man loves a woman who wrote song when a man loves a woman new england patriots who has participated in the most super b owls what nfl team has been to most super bowls

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

34

new england patriots who has participated in the most super b owls what nfl team has been to most super bowls Table 5: Additional examples of test-train overlapping que stions in Open Natural Questions Answer Test Question Train Question Picasso Who painted ”Boy With a Pipe” which, in May 2004, was sold for a record price of $104 million?painted in 1905, the painting garcon a la pipe was a famous painting by which famous artist who died in 1973? Wensum On what river is the city of Norwich the english city of norwich lies on which river? Mantle Comprising around two-thirds of the Earth’s mass , what is found between the core of the Earth and its crust?what do we call the layer of the earth between its crust and its core? Live and Let Die In which James Bond film does actress Jane Seym our play Solitaire?jane seymour played the character ”solitaire” in which bond film? Esau Who, in the Bible, was the eldest son of Isaac? in the bibl e, who was the first born of isaac? Alanis Morrisette Who made the 1995 album ’Jagged Little Pil l’ which

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

35

Alanis Morrisette Who made the 1995 album ’Jagged Little Pil l’ which sold 33 million copies?who released the 1995 hit album ”jagged little pill”? Excalibur In British legend, what is the name of King Arthurs sword?what was the name of king arthur’s sword? Humidity What is measured by a Hygrometer? what does a hygrom eter measure? A Storm On the Beaufort scale what is defined as force 11? what i s force 11 (eleven) on the beaufort scale? Jeremy Irons Actress Sinead Cusack is married to which ’Osca r’ winning actor?which actor is the husband of sinead cusack? Sir Cloudesley Shovell Who was the British Admiral who died i n 1707 when four of his ships were wrecked in the Scilly Isles?in 1707 a fleet of navy ships was wrecked off the scilly islands. who was the commander who lost his life in the disaster? Tony Hart Which famous individual created the ’Blue Peter’ s ailing ship logo?which artist designed the logo for uk television childrens show blue peter? Table 6: Examples of test-train overlapping questions in Tr iviaQA Answer Test Question Train Question

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

36

show blue peter? Table 6: Examples of test-train overlapping questions in Tr iviaQA Answer Test Question Train Question costa rica where is isthmus of panama located on the map? wher e is isthmus of panama located? 1986 world series when’s the last time the mets won the world s eries? when did the mets win the pennant? abbottabad where was bin laden found and killed? what countr y was osama bin laden killed in? believer what other movies has ryan gosling been in? what mov ies does ryan gosling star in? sculpture what type of art did leonardo da vinci make? what ki nd of art did leonardo da vinci produce? origin of species what book did charles darwin wrote in 1859? what was the name of the book that charles darwin wrote? morehouse college what college did martin luther king jr go t o? where did dr. martin luther king jr. go to school? communist state what type of government did soviet union hav e? what type of government does the former soviet union have? turkish lira what money to take to turkey? what currency to ta ke to side turkey?

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2008.02637

37

turkish lira what money to take to turkey? what currency to ta ke to side turkey? spanish language what is the most common language spoken in a rgentina?what is language in argentina? opera OR classical musicwhat music period did beethoven live in? what music did beeth oven composed? harry s truman who was president after franklin d. roosevelt ? who became president when roosevelt died in office? Table 7: Examples of test-train overlapping questions in We bQuestions

2008.02637

Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 60-70% of test-time answers are also present somewhere in the training sets. We also find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding training sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can actually generalize, and what drives their overall performance. We find that all models perform dramatically worse on questions that cannot be memorized from training sets, with a mean absolute performance difference of 63% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models out-perform a BART closed-book QA model, further highlighting the role that training set memorization plays in these benchmarks

http://arxiv.org/pdf/2008.02637

cs.CL

20200806

20200806

2108.10934

0

Mitigating Statistical Bias within Differentially Private Synthetic Data Sahra Ghalebikesabi1Harrison Wilde2Jack Jewson3Arnaud Doucet1 Sebastian Vollmer5Chris Holmes1 1University of Oxford 2University of Warwick 3Universitat Pompeu Fabra 5University of Kaiserslautern, German Research Centre for Artificial Intelligence (DFKI) Abstract Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacycompliant augmentation for general applications of synthetic data. 1 INTRODUCTION The prevalence of sensitive datasets, such as electronic health records, contributes to a growing concern for violations of an individual’s privacy. In recent years, the notion of

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519