Datasets:

jamescalam

/

ai-arxiv-chunked

like 32

2211.00053

45

Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. CommonGen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020 , pp. 1823–1840, Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.findings-emnlp.165. URL https://aclanthology.org/2020. findings-emnlp.165 . Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A. Smith, and Yejin Choi. DExperts: Decoding-time controlled text generation with experts and antiexperts. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pp. 6691–6706, Online, August 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.522. URL https://aclanthology.org/2021. acl-long.522 .

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

46

acl-long.522 . Xianggen Liu, Lili Mou, Fandong Meng, Hao Zhou, Jie Zhou, and Sen Song. Unsupervised paraphrasing by simulated annealing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pp. 302–312, Online, July 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.28. URL https://aclanthology.org/2020. acl-main.28 . Ximing Lu, Peter West, Rowan Zellers, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. NeuroLogic decoding: (un)supervised neural text generation with predicate logic constraints. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pp. 4288–4299, Online, June 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.naacl-main.339. URL https://aclanthology.org/2021.naacl-main.339 .

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

47

https://aclanthology.org/2021.naacl-main.339 . Ximing Lu, Sean Welleck, Liwei Jiang, Jack Hessel, Lianhui Qin, Peter West, Prithviraj Ammanabrolu, and Yejin Choi. Quark: Controllable text generation with reinforced unlearning. 12 Preprint. Under review. CoRR , abs/2205.13636, 2022a. doi: 10.48550/arXiv.2205.13636. URL https://doi.org/ 10.48550/arXiv.2205.13636 . Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, and Yejin Choi. NeuroLogic a*esque decoding: Constrained text generation with lookahead heuristics. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pp. 780–799, Seattle, United States, July 2022b. Association for Computational Linguistics. doi: 10.18653/v1/2022.naacl-main.57. URL https:

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

48

//aclanthology.org/2022.naacl-main.57 . Kathleen McKeown. Text Generation . Studies in Natural Language Processing. Cambridge University Press, 1985. doi: 10.1017/CBO9780511620751. Ning Miao, Hao Zhou, Lili Mou, Rui Yan, and Lei Li. Cgmh: Constrained sentence generation by metropolis-hastings sampling. Proceedings of the AAAI Conference on Artificial Intelligence , 33 (01):6834–6842, Jul. 2019. doi: 10.1609/aaai.v33i01.33016834. URL https://ojs.aaai. org/index.php/AAAI/article/view/4659 . Swaroop Mishra, Matthew Finlayson, Pan Lu, Leonard Tang, Sean Welleck, Chitta Baral, Tanmay Rajpurohit, Oyvind Tafjord, Ashish Sabharwal, Peter Clark, and Ashwin Kalyan. Lila: A unified benchmark for mathematical reasoning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , 2022.

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

49

benchmark for mathematical reasoning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , 2022. Ansong Ni, Jeevana Priya Inala, Chenglong Wang, Oleksandr Polozov, Christopher Meek, Dragomir Radev, and Jianfeng Gao. Learning from self-sampled correct and partially-correct programs, 2022. URL https://arxiv.org/abs/2205.14318 . Jekaterina Novikova, Ond ˇrej Dušek, and Verena Rieser. The E2E dataset: New challenges for end-to-end generation. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue , pp. 201–206, Saarbrücken, Germany, August 2017. Association for Computational Linguistics. doi: 10.18653/v1/W17-5525. URL https://aclanthology.org/W17-5525 . Arkil Patel, Satwik Bhattamishra, and Navin Goyal. Are NLP models really able to solve simple math word problems? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pp. 2080–

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

50

math word problems? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pp. 2080– 2094, Online, June 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021. naacl-main.168. URL https://aclanthology.org/2021.naacl-main.168 . Lianhui Qin, Sean Welleck, Daniel Khashabi, and Yejin Choi. Cold decoding: Energy-based constrained text generation with langevin dynamics. arXiv preprint arXiv:2202.11705 , 2022. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog , 1(8):9, 2019. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-totext transformer. Journal of Machine Learning Research , 21(140):1–67, 2020. URL http:

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

51

//jmlr.org/papers/v21/20-074.html . Machel Reid and Graham Neubig. Learning to model editing processes, 2022. URL https: //openreview.net/forum?id=1bEaEzGwfhP . Subhro Roy, Tim Vieira, and Dan Roth. Reasoning about quantities in natural language. Transactions of the Association for Computational Linguistics , 3:1–13, 2015. William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike. Self-critiquing models for assisting human evaluators, 2022. URL https://arxiv. org/abs/2206.05802 . Timo Schick, Jane Dwivedi-Yu, Zhengbao Jiang, Fabio Petroni, Patrick Lewis, Gautier Izacard, Qingfei You, Christoforos Nalmpantis, Edouard Grave, and Sebastian Riedel. Peer: A collaborative language model, 2022. URL https://arxiv.org/abs/2208.11663 . 13 Preprint. Under review. Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea V oss, Alec

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

52

13 Preprint. Under review. Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (eds.), Advances in Neural Information Processing Systems , volume 33, pp. 3008–3021. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/file/ 1f89885d556929e98d3ef9b86448f951-Paper.pdf . Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv , abs/2201.11903, 2022. Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

53

Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593 , 2019. URL https://arxiv.org/abs/1909.08593 . 14 Preprint. Under review. APPENDIX A A DDITIONAL EXPERIMENTAL DETAILS A.1 M ATHEMATICAL PROGRAM SYNTHESIS We fine-tune a separate instance of GPT-Neo 1.3B as an initial generator, using the Huggingface library with default hyperparameters, except for evaluation steps, which we set to a small number to ensure a strong checkpoint is selected for each dataset. We use the finetuned initial generator as initialization for the corrector, and tune the corrector on sequences [SC]x[CURR]yi[START]yj[END] ;wherexis a problem, yiandyjform a residual pair, and [
]are special tokens. The loss is on tokens after [START] . Feedback. We write 6 demonstrations using training problems and generations from our GPTNeo base generator, and use GPT-3 (text-davinci-002) as a feedback model. We use the same training procedure and hyperparameters, except that the sequences now include feedback,

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

54

same training procedure and hyperparameters, except that the sequences now include feedback, [SC]x[CURR]yi[FEEDBACK]F(x,yi)[START]yj[END] ;wherexis a problem, yiandyjform a residual pair, and F(x;yi)is feedback. We include loss on tokens after [FEEDBACK] . A.2 L EXICALLY -CONSTRAINED GENERATION Hyper-parameters. Table 8 and Table 9 show hyperparameters for CommonGen and E2E. Human Evaluation. We evaluate fluency of generations in E2E task using human annotators on Amazon Mechanical Turk (AMT). We randomly sampled 100 instances, along with generations of different baselines and self-corrections. For each instance, we ask 3 annotators to evaluate the fluency of generations on a 3-point Likert scale. We aggregate annotations from 3 annotators using majority vote. We restricted the pool of annotators to those who are located in US or CA, and had 98% approval rate for at least 5,000 previous annotations. Hyperparameter Assignment Predictor GPT-2 Large # steps 6000 batch size 128 optimizer Adam learning rate 1:e5 decoding alg. beam search (k=5)

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

55

# steps 6000 batch size 128 optimizer Adam learning rate 1:e5 decoding alg. beam search (k=5) Table 8: Hyperparameters for C OMMON GEN.Hyperparameter Assignment Predictor GPT-2 Medium # steps 10000 batch size 100 optimizer Adam learning rate 1:e5 decoding alg. beam search (k=5) Table 9: Hyperparameters for E2E. B A DDITIONAL RESULTS Toxicity Fluency Diversity Avg. Max .Prob. Perplexity dist-2 dist-3 GPT2-L 0.527 0.520 11.31 0.85 0.85 SELF-CORRECT 0.171 0.026 11.81 0.80 0.83 SELF-CORRECT +FEEDBACK 0.156 0.020 11.86 0.80 0.83 Table 10: Evaluation results of toxicity reduction experiments with natural language feedback. 15 Preprint. Under review. Bleu-4 CIDER Coverage Runtime NeuroLogic [22] 26.70 14.70 97.70 2.04s/sent NeuroLogic-A*esque [24] 28.20 15.20 97.80 19.24s/sent

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

56

NeuroLogic-A*esque [24] 28.20 15.20 97.80 19.24s/sent GPT-2 27.90 14.97 91.38 0.2s/sent SELF-CORRECT 27.98 15.30 94.58 0.8s/sent SELF-CORRECT + feedback 27.82 15.24 95.88 0.8s/sent SELF-CORRECT +NeuroLogic 28.17 15.28 97.80 2.24s/sent Table 11: Evaluation rresults of lexically-constrained generation on C OMMON GEN. Coverage BLEU-4 NIST R-L METEOR CIDER PREFIX -TUNING (Li & Liang, 2021) 91.16 70.30 8.82 72.10 46.30 2.46 GPT-2 91.50 67.12 8.67 70.25 45.58 2.33 SELF-CORRECT 98.77 68.81 8.78 68.60 45.11 2.38 Table 12: Evaluation results of lexically-constrained generation on E2E. C Q UALITATIVE EXAMPLES 16 Preprint. Under review. Problem: Ralph watches TV for 4 hours a day from

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

57

C Q UALITATIVE EXAMPLES 16 Preprint. Under review. Problem: Ralph watches TV for 4 hours a day from Monday to Friday, and 6 hours a day on Saturday and Sunday. How many hours does Ralph spend watching TV in one week? Generator: a=4*7 b=6*7 c=a+b answer=c print (answer)Corrector: a=4*5 b=6*2 c=a+b answer=c print (answer)Problem: The pirates plan to explore 4 islands. Two islands require walking 20 miles per day while the other two islands require 25 miles per day. How many miles will they have to walk if it takes 1.5 days to explore each island? Generator: a=20*2 b=25*2 c=a+b d=c*1.5 e=d+b answer=e print (answer)Corrector: a=20*2 b=25*2 c=a+b d=c*1.5 answer=d print (answer)

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

58

a=20*2 b=25*2 c=a+b d=c*1.5 answer=d print (answer) Figure 7: Grade school math (GSM) self-corrections. Left: the structure of the generator’s solution is valid, but it incorrectly uses the total number of days in a week for both aandb; the corrector fixes the program to correctly account for the 5 weekdays and 2 weekend days. Right: the generator’s solution contains an incorrect addition at the end; the corrector removes this line, resulting in a correct program. Problem: A spiral notebook costs 15, and a personal planner costs $10. How much would it cost in total to buy 4 spiral notebooks and 8 personal planners at a 20% discount? Generator: a=4*15 b=8*10 c=a+b answer=c print (answer)Corrector: a=4*15 b=8*10 c=a+b d=c*20/100 e=c-d answer=e print (answer)Problem: Julia has $40. She spends half of her money to buy a new game for her phone. She spends

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

59

answer=e print (answer)Problem: Julia has $40. She spends half of her money to buy a new game for her phone. She spends a quarter of what she has left on in-game purchases. How much money does she have left? Generator: a=40/2 b=40-a c=b/4 d=b-c e=d+40 answer=e print (answer)Corrector: a=40/2 b=a/4 c=a-b answer= c print (answer) Figure 8: Grade school math (GSM) self-corrections. Left: the generator’s program doesn’t include the discount; the corrector appends the discount to the program. Right: a more sophisticated multipart correction. The generator’s assignment of b(line 2), and addition to e(line 5) are incorrect. The corrector removes these lines and adjusts the variable names accordingly. 17 Preprint. Under review. Constraints: ‘table’, ‘paper’, ‘read’ Generator: A man is reading book on a table. Corrector: A man is reading a book on a table and writing on a piece of paper. Constraints:

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

60

Generator: A man is reading book on a table. Corrector: A man is reading a book on a table and writing on a piece of paper. Constraints: ’stand’, ’wait’, ’look’ Generator: He waits for someone to look at him. Corrector: He waits for someone to stand and look him in the eye. Constraints: ’walk’, ’park’, ’bench’, ’dog’, ’sit’ Generator: A dog sitting on a bench in a park. Corrector: A dog sitting on a bench and a man walking a dog in a park. Constraints: name : The Mill | Type : restaurant | food : English | price : high | customer rating : average | area : riverside | family friendly : yes | near : Cafe Rouge Generator: The Mill is an expensive, three star restaurant located near Cafe Rouge. Corrector: The Mill is a children friendly English restaurant in the riverside area near Cafe Rouge. It has an average customer rating and a high price range. Constraints: name : Blue Spice | Type : restaurant | food : Chinese | area : riverside | family friendly : yes | near : Rainbow Vegetarian Cafe Generator:

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

2211.00053

61

Constraints: name : Blue Spice | Type : restaurant | food : Chinese | area : riverside | family friendly : yes | near : Rainbow Vegetarian Cafe Generator: Blue Spice is a restaurant providing Chinese food. It is located in the riverside. It is near Rainbow Vegetarian Cafe. Corrector: Blue Spice is a family friendly Chinese restaurant in the riverside area near Rainbow Vegetarian Cafe. Figure 9: Constrained generation self-corrections. Top: C OMMON GEN, and Bottom: E2E. Words in green indicate constraints that were previously missing. Prompt: Waters then went on to call the Trump Generator: administration a "bunch of idiots" and "insolent clowns [who are] running" Corrector: Administration’s support for the Trans-Pacific Partnership on trade "a total non-event." Figure 10: Toxicity control self-corrections. Words in red indicate toxic language. 18

2211.00053

Generating Sequences by Learning to Self-Correct

Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct, using certain keywords, or avoiding undesirable content. Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs. Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation. We present Self-Correction, an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations. To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations. We show that Self-Correction improves upon the base generator in three diverse generation tasks - mathematical program synthesis, lexically-constrained generation, and toxicity control - even when the corrector is much smaller than the base generator.

http://arxiv.org/pdf/2211.00053

cs.CL

20221031

20221031

1906.07337

0

arXiv:1906.07337v1 [cs.CL] 18 Jun 2019Measuring Bias in Contextualized Word Representations Keita Kurita Nidhi Vyas Ayush Pareek Alan W Black Yulia Tsvet kov Carnegie Mellon University {kkurita,nkvyas,apareek,awb,ytsvetko }@andrew.cmu.edu Abstract Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1) propose a template-based method to quantify bias in BERT; (2) show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3) conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases. 1 Introduction Type-level word embedding models, including word2vec and GloVe ( Mikolov et al. ,2013 ;

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

1

such as racial and religious biases. 1 Introduction Type-level word embedding models, including word2vec and GloVe ( Mikolov et al. ,2013 ; Pennington et al. ,2014 ), have been shown to exhibit social biases present in humangenerated training data ( Bolukbasi et al. , 2016 ;Caliskan et al. ,2017 ;Garg et al. ,2018 ; Manzini et al. ,2019 ). These embeddings are then used in a plethora of downstream applications, which perpetuate and further amplify stereotypes (Zhao et al. ,2017 ;Leino et al. ,2019 ). To reveal and quantify corpus-level biases is word embeddings, Bolukbasi et al. (2016 ) used the word analogy task ( Mikolov et al. ,2013 ). For example, they showed that gendered male word embeddings likehe, man are associated with higher-status jobs like computer programmer anddoctor , whereas gendered words like sheorwoman are associated with homemaker andnurse . Contextual word embedding models, such as ELMo and BERT ( Peters et al. ,2018 ; Devlin et al. ,2019 ) have become increasingly common, replacing traditional type-levelembeddings and attaining new state of the art

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

2

Devlin et al. ,2019 ) have become increasingly common, replacing traditional type-levelembeddings and attaining new state of the art results in the majority of NLP tasks. In these models, every word has a different embedding, depending on the context and the language model state; in these settings, the analogy task used to reveal biases in uncontextualized embeddings is not applicable. Recently, May et al. (2019 ) showed that traditional cosine-based methods for exposing bias in sentence embeddings fail to produce consistent results for embeddings generated using contextual methods. We find similar inconsistent results with cosine-based methods of exposing bias; this is a motivation to the development of a novel bias test that we propose. In this work, we propose a new method to quantify bias in BERT embeddings ( §2). Since BERT embeddings use a masked language modelling objective, we directly query the model to measure the bias for a particular token. More specifically, we create simple template sentences containing the attribute word for which we want to measure bias (e.g. programmer ) and the target for bias (e.g. she for gender). We then mask the attribute and target tokens sequentially, to get a relative measure of

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

3

for gender). We then mask the attribute and target tokens sequentially, to get a relative measure of bias across target classes (e.g. male and female). Contextualized word embeddings for a given token change based on its context, so such an approach allows us measure the bias for similar categories divergent by the the target attribute ( §2). We compare our approach with the cosine similaritybased approach ( §3) and show that our measure of bias is more consistent with human biases and is sensitive to a wide range of biases in the model using various stimuli presented in Caliskan et al. (2017 ). Next, we investigate the effect of a specific type of bias in a specific downstream task: gender bias in BERT and its effect on the task of Gendered Pronoun Resolution (GPR) ( Webster et al. , 2018 ). We show that the bias in GPR is highly correlated with our measure of bias ( §4). Finally, we highlight the potential negative impacts of using BERT in downstream real world applications (§5). The code and data used in this work are publicly available.1 2 Quantifying Bias in BERT

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

4

(§5). The code and data used in this work are publicly available.1 2 Quantifying Bias in BERT BERT is trained using a masked language modelling objective i.e. to predict masked tokens, denoted as [MASK], in a sentence given the entire context. We use the predictions for these [MASK] tokens to measure the bias encoded in the actual representations. We directly query the underlying masked language model in BERT2to compute the association between certain targets (e.g., gendered words) andattributes (e.g. career-related words). For example, to compute the association between the target male gender and the attribute programmer , we feed in the masked sentence “[MASK] is a programmer” to BERT, and compute the probability assigned to the sentence ‘ heis a programmer” (ptgt). To measure the association, however, we need to measure how much more BERT prefers the male gender association with the attribute programmer , compared to the female gender. We thus re-weight this likelihood ptgtusing the prior bias of the model towards predicting the male gender. To do this, we mask out the attribute programmer and query BERT with the sentence “[MASK] is a

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

5

of the model towards predicting the male gender. To do this, we mask out the attribute programmer and query BERT with the sentence “[MASK] is a [MASK]”, then compute the probability BERT assigns to the sentence ‘ heis a [MASK]” ( pprior). Intuitively, pprior represents how likely the word heis in BERT, given the sentence structure and no other evidence. Finally, the difference between the normalized predictions for the words heandshe can be used to measure the gender bias in BERT for the programmer attribute. Generalizing, we use the following procedure to compute the association between a target and an attribute: 1. Prepare a template sentence e.g.“[TARGET] is a [ATTRIBUTE]” 2. Replace [TARGET] with [MASK] and computeptgt=P([MASK]=[TARGET] |sentence) 3. Replace both [TARGET] and [ATTRIBUTE] with [MASK], and compute prior probability pprior=P([MASK]=[TARGET] |sentence) 1https://bit.ly/2EkJwh1 2For all experiments we use the uncased version of BERT BASE

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

6

1https://bit.ly/2EkJwh1 2For all experiments we use the uncased version of BERT BASE https://storage.googleapis.com/bert_models/2018_10_ 18/uncased_L-12_H-768_A-12.zip .4. Compute the association as logptgt pprior We refer to this normalized measure of association as the increased log probability score and the difference between the increased log probability scores for two targets (e.g. he/she) as log probability bias score which we use as measure of bias. Although this approach requires one to construct a template sentence, these templates are merely simple sentences containing attribute words of interest, and can be shared across multiple targets and attributes. Further, the flexibility to use such templates can potentially help measure more finegrained notions of bias in the model. In the next section, we show that our proposed log probability bias score method is more effective at exposing bias than traditional cosine-based measures. 3 Correlation with Human Biases We investigate the correlation between our measure of bias and human biases. To do this, we apply the log probability bias score to the same set of attributes that were shown to exhibit human bias in experiments that were performed using

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

7

apply the log probability bias score to the same set of attributes that were shown to exhibit human bias in experiments that were performed using the Implicit Association Test ( Greenwald et al. , 1998 ). Specifically, we use the stimuli used in the Word Embedding Association Test (WEAT) (Caliskan et al. ,2017 ). Word Embedding Association Test (WEAT) : The WEAT method compares set of target concepts (e.g. male and female words) denoted as X andY(each of equal size N), with a set of attributes to measure bias over social attributes and roles (e.g. career/family words) denoted as Aand B. The degree of bias for each target concept tis calculated as follows: s(t,A,B) = [ meana∈Asim(t,a)−meanb∈Bsim(t,b)], where simis the cosine similarity between the embeddings. The test statistics is S(X,Y,A,B ) = [ meanx∈Xs(x,A,B)− meany∈Ys(y,A,B)], where the test is a permutation test over XandY. Thep-value is computed as

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

8

meany∈Ys(y,A,B)], where the test is a permutation test over XandY. Thep-value is computed as p= Pr[S(Xi,Yi,A,B)> S(X,Y,A,B )] The effect size is measured as d=S(X,Y,A,B ) stdt∈X∪Ys(t,A,B) Category Templates Pleasant/Unpleasant (Insects/Flowers) T are A, T is A Pleasant/Unpleasant (EA/AA) T are A, T is A Career/Family (Male/Female) T likes A, T like A, T is interes ted in A Math/Arts (Male/Female) T likes A, T like A, T is interested i n A Science/Arts (Male/Female) T likes A, T like A, T is interest ed in A Table 1: Template sentences used for the WEAT tests (T: targe t, A: attribute) Category Targets Templates Pleasant/Unpleasant (Insects/Flowers) flowers,insects, flower,insect T are A, the T is A

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

9

Pleasant/Unpleasant (Insects/Flowers) flowers,insects, flower,insect T are A, the T is A Pleasant/Unpleasant (EA/AA) black, white T people are A, th e T person is A Career/Family (Male/Female) he,she,boys,girls,men,wom en T likes A, T like A, T is interested in A Math/Arts (Male/Female) he,she,boys,girls,men,women T l ikes A, T like A, T is interested in A Science/Arts (Male/Female) he,she,boys,girls,men,wome n T likes A, T like A, T is interested in A Table 2: Template sentences used and target words for the gra mmatically correct sentences (T: target, A: attribute) It is important to note that the statistical test is a permutation test, and hence a large effect size does not guarantee a higher degree of statistical significance. 3.1 Baseline: WEAT for BERT To apply the WEAT method on BERT, we first compute the embeddings for target and attribute

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

10

3.1 Baseline: WEAT for BERT To apply the WEAT method on BERT, we first compute the embeddings for target and attribute words present in the stimuli using multiple templates, such as “TARGET is ATTRIBUTE” (Refer Table 1for an exhaustive list of templates used for each category). We mask the TARGET to compute the embedding3for the ATTRIBUTE and vice versa. Words that are absent in the BERT vocabulary are removed from the targets. We ensure that the number of words for both targets are equal, by removing random words from the smaller target set. To confirm whether the reduction in vocabulary results in a change of p-value, we also conduct the WEAT on GloVe with the reduced vocabulary.4 3.2 Proposed: Log Probability Bias Score To compare our method measuring bias, and to test for human-like biases in BERT, we also compute the log probability bias score for the same set of attributes and targets in the stimuli. We compute the mean log probability bias score for each attribute, and permute the attributes to measure statistical significance with the permutation test. Since many TARGETs in the stimuli cause the template sentence to become grammatically

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

11

test. Since many TARGETs in the stimuli cause the template sentence to become grammatically 3We use the outputs from the final layer of BERT as embeddings 4WEAT was originally used to study the GloVe embeddingsincorrect, resulting in low predicted probabilities, we fixed the TARGET to common pronouns/indicators of category such as flower, he, she(Table 2contains a full list of target words and templates). This avoids large variance in predicted probabilities, leading to more reliable results. The effect size is computed in the same way as the WEAT except the standard deviation is computed over the mean log probability bias scores . We experiment over the following categories of stimuli in the WEAT experiments: Category 1 (flower/insect targets and pleasant/unpleasant attributes), Category 3 (European American/African American names and pleasant/unpleasant attributes), Category 6 (male/female names and career/family attributes), Category 7 (male/female targets and math/arts attributes) and Category 8 (male/female targets and science/arts attributes). 3.3 Comparison Results The WEAT on GloVe returns similar findings to those of Caliskan et al. (2017 ) except for the

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

12

3.3 Comparison Results The WEAT on GloVe returns similar findings to those of Caliskan et al. (2017 ) except for the European/African American names and pleasant/unpleasant association not exhibiting significant bias. This is due to only 5 of the African American names being present in the BERT vocabulary. The WEAT for BERT fails to find any statistically significant biases at p <0.01. This implies that WEAT is not an effective measure for bias in BERT embeddings, or that methods for constructing embeddings require additional investigation. In contrast, our method of querying the underlying language model exposes statistically significant association across all categories, showing that BERT does indeed encode biases and that our method is more sensitive to them. Category WEAT on GloVe WEAT on BERT Ours on BERT Log Probability Bias Scor e Pleasant/Unpleasant (Insects/Flowers) 1.543* 0.6688 0.8 744* Pleasant/Unpleasant (EA/AA) 1.012 1.003 0.8864* Career/Family (Male/Female) 1.814* 0.5047 1.126*

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

13

Career/Family (Male/Female) 1.814* 0.5047 1.126* Math/Arts (Male/Female) 1.061 0.6755 0.8495* Science/Arts (Male/Female) 1.246* 0.8815 0.9572* Table 3: Effect sizes of bias measurements on WEAT Stimuli. ( * indicates significant at p <0.01) Gender Prior Prob. Avg. Predicted Prob. Male 10.3% 11.5% Female 9.8% 13.9% Table 4: Probability of pronoun referring to neither entity in a sentence of GPR 4 Case Study: Effects of Gender Bias on Gendered Pronoun Resolution Dataset We examined the downstream effects of bias in BERT using the Gendered Pronoun Resolution (GPR) task ( Webster et al. ,2018 ). GPR is a sub-task in co-reference resolution, where a pronoun-containing expression is to be paired with the referring expression. Since pronoun resolving systems generally favor the male entities (Webster et al. ,2018 ), this task is a valid testbed for our study. We use the GAP dataset5 byWebster et al. (2018 ), containing 8,908 humanlabeled ambiguous pronoun-name pairs, created

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

14

byWebster et al. (2018 ), containing 8,908 humanlabeled ambiguous pronoun-name pairs, created from Wikipedia. The task is to classify whether an ambiguous pronoun Pin a text refers to entity A, entityBor neither. There are 1,000 male and female pronouns in the training set each, with 103 and 98 of them not referring to any entity in the sentence, respectively. Model We use the model suggested on Kaggle,6 inspired by Tenney et al. (2019 ). The model uses BERT embeddings for P,AandB, given the context of the input sentence. Next, it uses a multilayer perceptron (MLP) layer to perform a naive classification to decide if the pronoun belongs to A,Bor neither. The MLP layer uses a single hidden layer with 31 dimensions, a dropout of 0.6 and L2 regularization with weight 0.1. Results Although the number of male pronouns associated with no entities in the training data is slightly larger, the model predicted the female pro5https://github.com/google-research-datasets/gap-cor eference 6https://www.kaggle.com/mateiionita/taming-the-bert- a-baselinenoun referring to no entities with a significantly

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

15

higher probability ( p= 0.007 on a permutation test); see Table 4. As the training set is balanced, we attribute this bias to the underlying BERT representations. We also investigate the relation between the topic of the sentence and model’s ability to associate the female pronoun with no entity. We first extracted 20 major topics from the dataset using non-negative matrix factorization ( Lee and Seung , 2001 ) (refer to Appendix for the list of topics). We then compute the bias score for each topic as the sum of the log probability bias score for the top 15 most prevalent words of each topic weighted by their weights within the topic. For this, we use a generic template “[TARGET] are interested in [ATTRIBUTE]” where TARGET is either men or women. Next we compute a bias score for each sample in the training data as the sum of individual bias scores of topics present in the sample, weighted by the topic weights. Finally, we measured the Spearman correlation coefficient to be 0.207 (which is statistically significant with p= 4e−11) between the bias scores for male gender across all samples and the model’s probability to associate a female pronoun with no entity.

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

16

p= 4e−11) between the bias scores for male gender across all samples and the model’s probability to associate a female pronoun with no entity. We conclude that models using BERT find it challenging to perform coreference resolution when the gender pronoun is female and if the topic is biased towards the male gender. 5 Real World Implications In previous sections, we discussed that BERT has human-like biases, which are propagated to downstream tasks. In this section, we discuss another potential negative impact of using BERT in a downstream model. Given that three quarters of US employers now use social media for recruiting job candidates ( Segal ,2014 ), many applications are filtered using job recommendation systems and other AI-powered services. Zhao et al. (2018 ) discussed that resume filtering systems are biased when the model has strong association between gender and certain professions. Similarly, certain gender-stereotyped attributes have been strongly associated with occupational salary and prestige (Glick ,1991 ). Using our proposed method, we investigate the gender bias in BERT embeddingss for certain occupation and skill attributes. Datasets : We use three datasets for our study of gender bias in employment attributes: •Employee Salary Dataset7for Montgomery

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

17

for certain occupation and skill attributes. Datasets : We use three datasets for our study of gender bias in employment attributes: •Employee Salary Dataset7for Montgomery County of Maryland- Contains 6882 instances of “Job Title” and “Salary” records along with other attributes. We sort this dataset in decreasing order of salary and take the first 1000 instances as a proxy for highpaying and prestigious jobs. •Positive and Negative Traits Dataset8-Contains a collection of 234 and 292 adjectives considered “positive” and “negative” traits, respectively. •O*NET 23.2 technology skills9Contains 17649 unique skills for 27660 jobs, which are posted online Discussion We used the following two templates to measure gender bias: •“TARGET is ATTRIBUTE”, where TARGET are male and female pronouns viz. he andshe. The ATTRIBUTE are job titles from the Employee Salary dataset, or the adjectives from the Positive and Negative traits dataset. •“TARGET can do ATTRIBUTE”, where the TARGETs are the same, but the ATTRIBUTE are skills from the O*NET dataset. Table 5shows the percentage of attributes that

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

18

the TARGETs are the same, but the ATTRIBUTE are skills from the O*NET dataset. Table 5shows the percentage of attributes that were more strongly associated with the male than the female gender. The results prove that BERT expresses strong preferences for male pronouns, raising concerns with using BERT in downstream tasks like resume filtering. 7https://catalog.data.gov/dataset/employee-salaries- 2017 8http://ideonomy.mit.edu/essays/traits.html 9https://www.onetcenter.org/database.html#individual -filesDataset Percentage Salary 88.5% Pos-Traits 80.0% Neg-Traits 78.9% Skills 84.0% Table 5: Percentage of attributes associated more strongly with the male gender 6 Related Work NLP applications ranging from core tasks such as coreference resolution ( Rudinger et al. ,2018 ) and language identification ( Jurgens et al. ,2017 ), to downstream systems such as automated essay scoring ( Amorim et al. ,2018 ), exhibit inherent social biases which are attributed to the datasets used to train the embeddings ( Barocas and Selbst ,

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

19

scoring ( Amorim et al. ,2018 ), exhibit inherent social biases which are attributed to the datasets used to train the embeddings ( Barocas and Selbst , 2016 ;Zhao et al. ,2017 ;Yao and Huang ,2017 ). There have been several efforts to investigate the amount of intrinsic bias within uncontextualized word embeddings in binary ( Bolukbasi et al. , 2016 ;Garg et al. ,2018 ;Swinger et al. ,2019 ) and multiclass ( Manzini et al. ,2019 ) settings. Contextualized embeddings such as BERT (Devlin et al. ,2019 ) and ELMo ( Peters et al. , 2018 ) have been replacing the traditional typelevel embeddings. It is thus important to understand the effects of biases learned by these embedding models on downstream tasks. However, it is not straightforward to use the existing biasexposure methods for contextualized embeddings. For instance, May et al. (2019 ) used WEAT on sentence embeddings of ELMo and BERT, but there was no clear indication of bias. Rather, they observed counterintuitive behavior like vastly differentp-values for results concerning gender. Along similar lines, Basta et al. (2019 ) noted that contextual word-embeddings are less biased

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

20

Along similar lines, Basta et al. (2019 ) noted that contextual word-embeddings are less biased than traditional word-embeddings. Yet, biases like gender are propagated heavily in downstream tasks. For instance, Zhao et al. (2019 ) showed that ELMo exhibits gender bias for certain professions. As a result, female entities are predicted less accurately than male entities for certain occupation words, in the coreference resolution task. Field and Tsvetkov (2019 ) revealed biases in ELMo embeddings that limit their applicability across data domains. Motivated by these recent findings, our work proposes a new method to expose and measure bias in contextualized word embeddings, specifically BERT. As opposed to previous work, our measure of bias is more consistent with human biases. We also study the effect of this intrinsic bias on downstream tasks, and highlight the negative impacts of gender-bias in real world applications. 7 Conclusion In this paper, we showed that querying the underlying language model can effectively measure bias in BERT and expose multiple stereotypes embedded in the model. We also showed that our measure of bias is more consistent with human-biases, and outperforms the traditional WEAT method on BERT. Finally we showed that these biases can

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

21

and outperforms the traditional WEAT method on BERT. Finally we showed that these biases can have negative downstream effects. In the future, we would like to explore the effects on other downstream tasks such as text classification, and device an effective method of debiasing contextualized word embeddings. Acknowledgments This material is based upon work supported by the National Science Foundation under Grant No. IIS1812327. References Evelin Amorim, Marcia Canc ¸ado, and Adriano Veloso. 2018. Automated essay scoring in the presence of biased ratings. In Proc. of NAACL , pages 229–237. Solon Barocas and Andrew D Selbst. 2016. Big data’s disparate impact. Calif. L. Rev. , 104:671. Christine Basta, Marta R Costa-juss` a, and Noe Casas. 2019. Evaluating the underlying gender bias in contextualized word embeddings. arXiv preprint arXiv:1904.08783 . Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

22

Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Proc. of NIPS , pages 4349–4357. Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science , 356(6334):183–186. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. of NAACL . Anjalie Field and Yulia Tsvetkov. 2019. Entity-centric contextual affective analysis. In Proc. of ACL .Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences , 115(16):E3635–E3644. Peter Glick. 1991. Trait-based and sex-based discrimination in occupational prestige, occupational salary, and hiring. Sex Roles , 25(5-6):351–378. Anthony Greenwald, Debbie E. McGhee, and Jordan

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

23

and hiring. Sex Roles , 25(5-6):351–378. Anthony Greenwald, Debbie E. McGhee, and Jordan L. K. Schwartz. 1998. Measuring individual differences in implicit cognition: The implicit association test. Journal of personality and social psychology , 74:1464–80. David Jurgens, Yulia Tsvetkov, and Dan Jurafsky. 2017. Incorporating dialectal variability for socially equitable language identification. In Proc. of ACL , pages 51–57. Daniel Lee and Hyunjune Seung. 2001. Algorithms for non-negative matrix factorization. In Proc. of NIPS . Klas Leino, Matt Fredrikson, Emily Black, Shayak Sen, and Anupam Datta. 2019. Feature-wise bias amplification. In Prof. of ICLR . Thomas Manzini, Yao Chong, Yulia Tsvetkov, and Alan W Black. 2019. Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. In Proc. of NAACL . Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. In Proc. of

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

24

Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. In Proc. of NAACL . Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Proc.of NIPS , pages 3111–3119. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proce. of EMNLP , pages 1532– 1543. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proc. of NAACL . Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Proc. of NAACL . J Segal. 2014. Social media use in hiring: Assessing the risks. HR Magazine , 59(9). Nathaniel Swinger, Maria De-Arteaga, Neil Heffernan IV , Mark Leiserson, and Adam Kalai. 2019. What are the biases in my word embedding? In

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

25

What are the biases in my word embedding? In Proc. of the AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society (AIES) . Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R. Thomas McCoy, Najoung Kim, Benjamin Van Durme, Samuel R. Bowman, Dipanjan Das, and Ellie Pavlick. 2019. What do you learn from context? probing for sentence structure in contextualized word representations. In Proc. of ICLR . 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 . Sirui Yao and Bert Huang. 2017. Beyond parity: Fairness objectives for collaborative filtering. In Advances in Neural Information Processing Systems , pages 2921–2930. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Ryan Cotterell, Vicente Ordonez, and Kai-Wei Chang. 2019. Gender bias in contextualized word embeddings. In NAACL (short) . Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

26

NAACL (short) . 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 Proc. of EMNLP . Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, and KaiWei Chang. 2018. Learning gender-neutral word embeddings.Appendix Topic Id Top 5 Words 1 match,round,second,team,season 2 times,city,jersey,york,new 3 married,son,died,wife,daughter 4 best,award,actress,films,film 5 friend,like,work,mother,life 6 university,music,attended,high,school 7 president,general,governor,party,state 8 songs,solo,song,band,album 9 medal,gold,final,won,world 10 best,role,character,television,series 11 kruse,moved,amy,esme,time 12 usa,trunchbull,pageant,2011,miss 13 american,august,brother,actress,born

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1906.07337

27

12 usa,trunchbull,pageant,2011,miss 13 american,august,brother,actress,born 14 sir,died,church,song,john 15 natasha,days,hospital,helene,later 16 played,debut,sang,role,opera 17 january,december,october,july,married 18 academy,member,american,university,family 19 award,best,played,mary,year 20 jersey,death,james,king,paul Table 6: Extracted topics for the GPR dataset

1906.07337

Measuring Bias in Contextualized Word Representations

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.

http://arxiv.org/pdf/1906.07337

1st ACL Workshop on Gender Bias for Natural Language Processing 2019

cs.CL

20190618

20190618

1907.11274

0

arXiv:1907.11274v2 [cs.CY] 29 Jul 2019REDUCING MALICIOUS USE OF SYNTHETIC MEDIA RESEARCH : CONSIDERATIONS AND POTENTIAL RELEASE PRACTICES FOR MACHINE LEARNING Aviv Ovadya The Thoughtful Technology Project aviv.ovadya@thoughtfultech.orgJess Whittlestone Leverhulme Centre for the Future of Intelligence University of Cambridge jlw84@cam.ac.uk ABSTRACT The aim of this paper is to facilitate nuanced discussion aro und research norms and practices to mitigate the harmful impacts of advances in machine learnin g (ML). We focus particularly on the use of ML to create “synthetic media” (e.g. to generate or man ipulate audio, video, images, and text), and the question of what publication and release proc esses around such research might look like, though many of the considerations discussed will appl y to ML research more broadly. We are not arguing for any specific approach on when or how research s hould be distributed, but instead try to lay out some useful tools, analogies, and options for thin king about these issues.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

1

to lay out some useful tools, analogies, and options for thin king about these issues. We begin with some background on the idea that ML research mig ht be misused in harmful ways, and why advances in synthetic media, in particular, are rais ing concerns. We then outline in more detail some of the different paths to harm from ML research, b efore reviewing research risk mitigation strategies in other fields and identifying component s that seem most worth emulating in the ML and synthetic media research communities. Next, we outli ne some important dimensions of disagreement on these issues which risk polarizing convers ations. Finally, we conclude with recommendations, suggesting tha t the machine learning community might benefit from: working with subject matter experts to increas e understanding of the risk landscape and possible mitigation strategies; building a community and n orms around understanding the impacts of ML research, e.g. through regular workshops at major conf erences; and establishing institutions and systems to support release practices that would otherwi se be onerous and error-prone. Keywords synthetic media ·research practices ·release practices ·societal impacts ·negative impacts ·machine learning·malicious use ·dual-use·deepfakes ·fake news 1 Introduction Technological advances can result in harm to both individua ls and societal structures, through accidents, unintended

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

2

1 Introduction Technological advances can result in harm to both individua ls and societal structures, through accidents, unintended consequences, and malicious use – even as the same advances p rovide incredible benefits. Concern about harms resulting from advances in machine learning (ML) have risen dramat ically in recent years, with a growing community of researchers and practitioners doing crucial work to addres s issues such as the fairness, accountability, and transpar ency of deployed systems [1]. In this paper, we focus primarily on the ways that advances in ML research might be deliberately misused by malicious actors (which we refer to as “ma l-use”) [2], though we also touch on other unintended consequences.1 1Unintended consequences are also particularly salient for the case of synthetic video and audio, where the simple exist ence of the technology, regardless of its use, can allow bad actors t o claim that evidence of e.g. corruption or war crimes were sy nthesized in order to avoid accountability. For example, allegations of videos being faked have been used to justify a coup in Gabon [3], and exculpate a cabinet minister in Malaysia [4]. DRAFT - JULY 2019

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

3

exculpate a cabinet minister in Malaysia [4]. DRAFT - JULY 2019 One specific example of where mal-use concerns have been rais ed recently is around the use of ML in synthetic media2: the manipulation and generation of increasingly realistic audio, video, images, and text. It is now possible to produce synthetic images of faces that are almost indistinguishabl e from photographs [5]. It is becoming easier to manipulate existing videos of people: replacing a person’s facial expr essions or movements with those of another person [6, 7], and/or overlaying synthesized speech which imitates a pers on’s voice [8]. Language models have advanced to the point where, given a sentence prompt, they can write article s that appear at least superficially convincing [9]. These advances could be used to impersonate people, sway public op inion, or more generally spread doubt about the veracity of all media. Modern synthetic media is in fact already being used for harm: face-swapping tools are being used to harass journalists [10], synthetic voices are being used fo r financial crimes [11], and synthetic faces have allegedly been used for espionage [12]. These examples bring into focus the need for humility around what we don’t know concerning potential impacts

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

4

been used for espionage [12]. These examples bring into focus the need for humility around what we don’t know concerning potential impacts of technology, and perhaps suggests that we must develop bet ter systems for understanding such impacts. Related concerns have sparked debate about responsible research pr actices in ML [13], and in particular whether sometimes it is appropriate to withhold open publication of some aspects of this research. We begin with some background on the idea that ML research mig ht be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then review research risk mitigation strategies in other fields and identifying components that may be worth emulatin g in the ML and synthetic media research communities. Finally, we outline some important dimensions of disagreem ent on these issues which risk polarizing conversations, before concluding with some recommendations for research a nd practice. 2 How research can lead to harm How might advances in ML and synthetic media end up resulting in harm? We begin by spelling this out in more detail: different ways that research might empower malicio us actors, and some possible paths to harm from this. 2.1 Types of hazard We can distinguish several different types of “information hazard” that might result from machine learning research in general.3

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

5

2.1 Types of hazard We can distinguish several different types of “information hazard” that might result from machine learning research in general.3 •Product hazard: Research produces software that can be directly used for har m(e.g. rootkits for computer hacking). Product hazards increase the likelihood of mal-u se by adversaries with minimal technical capabilities (e.g. ‘script kiddies’ who use existing programs or sc ripts to hack into computers but lack the ability to write their own, or less sophisticated information warfare operations), or those who have weak motivations for harm (e.g. online hacking or deception just ‘for fun’). •Data hazard: research produces detailed information or outputs which if disseminated, create risk of use for harm (e.g. easy nuclear blueprints; models, training data or cod e for harmful software). Data hazards increase the likelihood of mal-use by adversaries with some technical capabilities, but without e.g. access to high-quality researchers. •Attention hazard: research which directs attention towards an idea or data tha t increases risk (e.g. the idea that it is possible to use voice-cloning for phishing). Atte ntion hazards increase the likelihood of mal-use by adversaries who may not have realized that their objectives can be aided by new technologies.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

6

adversaries who may not have realized that their objectives can be aided by new technologies. An important thing to note is that potential mitigations wil l bedifferent for each type of hazard – and potentially even in conflict . One way to mitigate attention hazards is to be very careful a bout talking to media organizations and others with large reach about ways that research advances could be u sed maliciously (such as in voice cloning, for example). At the same time, raising wider concern about malicious use c ases of ML progress might be exactly what is needed to incentivize mitigations for data hazards or product haza rds (e.g. some public concern may be required to ensure that tech platforms prioritize developing technology to id entify voice cloning, or for telecoms to build mitigating infrastructure to make phone number spoofing harder). 2Not all synthetic media techniques involve ML. For example, traditional computer graphics techniques are also used for image and video manipulation. For simplicity of language, we will not consider these distinctly as the considerations for res earch practices are very similar. 3This is based loosely on a taxonomy from [14] 2 DRAFT - JULY 2019 2.2 The path to harm But how might these different types of ‘hazard’ actually lea d to real-world harms? It is worth connecting the dots here

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

7

2.2 The path to harm But how might these different types of ‘hazard’ actually lea d to real-world harms? It is worth connecting the dots here between the theoretical potential for mal-use, and what act ually makes significant harm more likely. Below we talk through some factors influencing whether a capa bility leads to sustained mal-use in practice. We use artificial voice cloning as an illustrative example, as a rel atively new capability with many useful applications (e.g. in voice translation and audio editing) but also significant po tential for mal-use (e.g. in scams, political propaganda, a nd market manipulation). 1.Awareness : Do actors with malicious intent know about a capability and believe it can help them? We can break this down into: •Attention of adversaries : Are malicious actors likely to realize that they could use a new capability to further their ends? If adversary groups are already using closely related methods, this is much more likely: for example, if edited voice clips are already being used for political manipulation, groups doing this are more likely to pay attention to demonstrations of vo ice cloning. •‘Convincibility’ of those with resources : Are there compelling arguments, perhaps by authoritative

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

8

this are more likely to pay attention to demonstrations of vo ice cloning. •‘Convincibility’ of those with resources : Are there compelling arguments, perhaps by authoritative third parties, for the effectiveness of new capabilities? F or example, a scammer who realizes that voice cloning is useful might need to be able to convince a superior that this technology is effective enough to justify the costs and overcome institutional inertia. 2.Deployment : How difficult is it for adversaries to weaponize this capabi lity in practice? For a capability to be deployed for malicious purposes, adve rsaries not only need to be aware but to have the necessary skills and resources to productize and weaponize the capability. This isn’t a binary – e.g. having ML expertise vs. not – but rather many different factors will influence how easy a capability is to weaponize. At the extreme, we might have a product which can be immediate ly used by anyone, regardless of technical capability (such as free to use voice cloning software). Factors that influence the ease of deployment for mal-use inc lude: •Talent pipelines : How difficult is it to source someone who can apply a new capab ility for the desired

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

9

•Talent pipelines : How difficult is it to source someone who can apply a new capab ility for the desired use case? (e.g. do malicious actors need someone with machin e learning experience, programming experience, or can they just use a program directly to achiev e their goals?) [12]. •Reproducibility : How difficult is it to reproduce a capability given the infor mation available? (e.g. is it easy to replicate a voice cloning capability given the avail able papers, models, code, etc.?) •Modifiability : How difficult is it to modify or use a system in order to enable mal-use? (e.g. if a voice cloning product makes it difficult to clone a voice without co nsent or watermarks, how hard is it to overcome those limitations?) •Slottability : Can new capabilities be slotted into existing organizatio nal processes or technical systems? (e.g. are there already established processes for phone sca ms into which new voice generation capabilities can be slotted easily, without any need to change goals o r strategy?) •Environmental factors : How does the existing ‘environment’ or ‘infrastructure’ i mpact the usefulness

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

10

•Environmental factors : How does the existing ‘environment’ or ‘infrastructure’ i mpact the usefulness of the new capability for malicious actors? (E.g. currently , in the US it is easy to ‘spoof’ phone numbers to make it appear like a call is coming from a family member, wh ich could impact the likelihood of voice cloning being weaponized for phone scams.) Websites now enabling anyone to instantly generate seeming ly photorealistic faces are a concrete example of deployment barriers falling away and making mal-use easi er. It had been possible for well over a year to generate synthetic images of faces with fairly high quality , but such websites have enabled anyone to do so with no technical expertise. This capability can also immed iately slot into existing processes, such as fake account creation. Previously, malicious actors would ofte n use existing photos of real people, which could be identified with reverse image search [15], unlike wholly gen erated synthetic images. 3.Sustained use : How likely is it that a capability will lead to sustained use with substantial negative impacts? Even if adversaries are aware of and able to weaponize some ne w capability, whether or not this leads to sustained use depends on:

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

11

Even if adversaries are aware of and able to weaponize some ne w capability, whether or not this leads to sustained use depends on: •Actual ROI : If malicious actors believe that the return on investment ( ROI) for using a capability is low they might not continue to use it in practice. For example, if a form of mal-use is easy to detect, then adversaries might decide it’s not worth the risk or might be s hut down very quickly. 3 DRAFT - JULY 2019 •Assessment of ROI : If malicious actors have no way of assessing whether new cap abilities are helping them better achieve their goals, or if their assessments are flawed, they might not continue to put resources into using those capabilities. 2.3 Access ratchets We can think of this as a kind of progression, from a theoretic al capability to scaled-up use in practice. Once a technology has progressed down this path and has become easy to use, and proven to have high ROI for mal-use, it can be much more difficult to address than at earlier stages – w e call this the access ratchet (like a ratchet, increased access to technology cannot generally be undone). For any ca pability with potential for mal-use, it is therefore worth

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

12

access to technology cannot generally be undone). For any ca pability with potential for mal-use, it is therefore worth thinking about where it currently sits on this progression: how much attention and interest it is receiving; whether it has been weaponized and/or how costly it would be to do so; and whether it’s likely to be, or already in sustained use. This can help us think more clearly about where the greatest r isks of mal-use are, and different kinds of interventions that might be appropriate or necessary in a given situation. Researchers may argue that a capability is unlikely to cause harm since it has not been used maliciously yet. What this doesn’t address is the fact that a capability which has n ot yet been used maliciously might sit anywhere along this progression, which makes a huge difference to how likely it i s to cause harm. For example, Face2Face, a technique for real-time facial reenactment (i.e. changing a person’s expressions in a video), has existed for 3 years but not been developed into any products that can easily be used. Thi s lack of productization makes harmful use vastly less likely, especially given the competition for AI and enginee ring talent today. It is also worth considering how costly it would be to make a given capability easier to misuse: even t he DeepFake application, which is more accessible to

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

13

it would be to make a given capability easier to misuse: even t he DeepFake application, which is more accessible to non-technical users, is currently resource-intensive to w eaponize in practice. 2.4 Indirect harms Sometimes the path to harm from synthetic media research wil l be fairly direct and immediate: such as a person losing their money, returning to our example of voice cloning being used in financial scams. But in other cases, improved synthetic media capabilities m ight cause harm in more complex and indirect ways. Consider the case where misinformation purveyors get hold of so phisticated synthetic media capabilities and use them to win substantial democratic power, which they then use to c ontrol narratives further and undermine any mitigation efforts (not an uncommon path from democracy to authoritari anism). We can think about this as a disinformation ratchet : the ability to use disinformation to enhance one’s ability to distribute further disinformation; and the opportunity for this type of ratchet can be influenced by new tec hnology impacting media distribution channels and capabilities. These less direct kinds of harms may be harder to anticipate o r imagine, but in the long-run may be much more important – particularly if they influence the future develo pment of technology in ways that undermine our ability to

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

14

important – particularly if they influence the future develo pment of technology in ways that undermine our ability to deal with future threats. We suggest that it’s particularly important to consider these kinds of “sociotechnical-path dependencies” as well as more direct and immediate threats, and what kinds of risk mitigation strategies might best address them. 3 Mitigating harm through release practices There is unlikely to be any ‘one size fits all’ solution to miti gating mal-use of ML research: the path to harm will look very different across contexts, and potential harms need to be weighed against benefits which will also vary depending on the area. We therefore need discussion about different ap proaches to mitigating mal-use: including around what research is conducted in the first place; standards and proce dures for risk assessment; and processes for deciding when and how to release different types of research outputs. Here we focus particularly on the latter – how careful release practices might help mitigate mal-use within ML research. However, this is not to suggest we think release practices ar e the main or even necessarily the most important component of mitigating mal-use. Another crucial piece is how r esearch directions are chosen and prioritized in the first

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

15

place. This is challenging because much of ML research often involves developing general capabilities which can then be applied to a variety of different purposes – we can’t simpl y decide to build only ‘beneficial’ ML capabilities. That aside, we still may be able to say some very general things abo ut the kinds of capabilities that are more likely to be broadly beneficial, or the kinds of problems that should idea lly be driving ML research. It is also important to think about what types of research are encouraged/discouraged by conferences, journals, funders, job interviewers and so on. 4 DRAFT - JULY 2019 3.1 Challenges to mitigating harm First, it’s worth considering some of the serious challenge s to attempting to decrease harm by limiting access to research: •Thecomposition problem : Two independent pieces of research that seem innocent can b e combined in ways that enable significant malicious use.4 •Theslow drip problem : Research advancement can be a slow and continuous evolutio n, where it’s difficult to draw the line between research that is dangerous and that w hich is not.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

16

to draw the line between research that is dangerous and that w hich is not. •Theconflation problem : Many of the underlying goals of various fields of research (n atural language processing, computation photography, etc.) may be directly we aponizable if achieved. For example, the ability to create convincing dialogue can be used to both support peo ple or manipulate people at scale. •The defector problem : Even if researchers in some regions or organizations coope ratively decide not to pursue or publish a particular area of research, those agree ments might not be followed by “defectors” who then gain a competitive edge. These challenges may seem daunting even for those who would a dvocate for limiting release of some forms of research. They also motivate the development of a nuanced menu of optio ns for release practices, and careful evaluation of the efficacy of whatever measures are chosen. Even without overc oming these challenges, it is possible that release practices could substantially mitigate harm if they impact the rate of deployment of mal-use technology.5 3.2 A brief tour of analogs in other fields

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

17

practices could substantially mitigate harm if they impact the rate of deployment of mal-use technology.5 3.2 A brief tour of analogs in other fields There is precedent in several other fields of research – inclu ding biotechnology and information security – for establishing processes for reducing the negative risks of resear ch and release. A good first step would, therefore, be to look at what we can learn from these fields for the case of ML res earch. Here we present some promising practices identified from other fields. A caveat: just because research norms and processes exist in other fields, it does not necessarily mean that they are widely and coherently used in those fields, or that they provi de a net positive impact. Evaluating which research practices have been adopted and work well across different fi elds is out of scope for this short paper, but would certainly be valuable to look into further. 3.2.1 Biosafety Biosafety processes and principles exist to ensure safe han dling of infective microorganisms in biology/biotechnolo gy research [16]. Some key components of biosafety practices i nclude:

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

18

research [16]. Some key components of biosafety practices i nclude: •Procedures : Steps and rules that must be followed, e.g. for decontamina tion (including basics such as wearing gloves and shoes). •Lab safety officer : An internal role responsible for enforcing safety. •Training : Learning safety processes via peers/programs. •Architecture : Incorporating safety considerations into building and to ol design (e.g. the design of doors and airflow). •Audits : Providing external accountability, usually at random tim es via local government. •Safety level designations : Different microorganisms classified by risk group (e.g. Eb ola is level 4) with different safety procedures for different levels (e.g. lev el 1 is open bench work, level 4 requires special clothing, airlock entry, special waste disposal, etc.). •Safety level definers : Organisations who determine safety levels, e.g. the Cente rs for Disease Control and Prevention (CDC). 4This might be particularly challenging given the success of transfer learning. 5In terms of the ratchet terminology used earlier, delaying r elease of research could slow down the speed of an ‘access rat chet’

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

19

5In terms of the ratchet terminology used earlier, delaying r elease of research could slow down the speed of an ‘access rat chet’ (i.e. slowing down widespread access to a technology), pote ntially providing enough extra time strengthen a ‘sociotec hnical immune system’ that could halt a disinformation ratchet. 5 DRAFT - JULY 2019 3.2.2 Computer/Information security Various practices exist in the field of information security to prevent exploitation of vulnerabilities in important sy stems. Key components include: •OPSEC (‘operations security’) : Procedures for identifying and protecting critical infor mation that could be used by adversaries. Includes identification of critical in formation, analysis of threats, vulnerabilities, and risks, and application of appropriate measures. •Architecture : Use systems that are “secure by design” and so keep you secur e automatically where possible. •Coordinated/responsible disclosure : Processes to ensure that exploits which could affect impor tant systems are not publicly disclosed until there has been an opportuni ty to fix the vulnerability. •ISACs/CERTs (Information Sharing & Analysis Centers/Computer Emergen cy Response Teams): Disclosure coordination entities.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

20

•ISACs/CERTs (Information Sharing & Analysis Centers/Computer Emergen cy Response Teams): Disclosure coordination entities. 3.2.3 Institutional Review Boards (IRBs) IRBs are designed to protect human subjects in biomedical an d behavioral research (including e.g. clinical trials of new drugs or devices and psychology studies of behavior, opi nions or attitudes) [17]: •Case dependent scrutiny : Research proposals are assessed on a case-by-case basis us ing external expert evaluation, and are determined to either be: (a) exempt (whe n risks are minimal), (b) expedited (slightly more than minimal risk), or (c) full review (all other propos als). •Approval rubrics : Criteria for approval of research proposals include: havi ng sound research principles to minimize risk to subjects; establishing that risks to subje cts are reasonable relative to anticipated benefits; selecting subjects in equitable ways, and avoiding undue em phasis on a vulnerable population. •External expert & community evaluation : Studies are reviewed by people who have expertise in the res earch and in the impacts of the work (such as community members). •Continuous evaluation : Process can be ongoing, not one-time, with periodic update s: the IRB can suspend or terminate previously approved research.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

21

•Continuous evaluation : Process can be ongoing, not one-time, with periodic update s: the IRB can suspend or terminate previously approved research. This is not meant to be exhaustive but demonstrates a variety of systems that have been used to mitigate negative risks of research and release. Other analogs worth explorin g include those around nuclear technology, spam detection, classified information, and environmental impact. 3.3 Potential release practices What should ML and synthetic media research emulate from the se other fields? Many aspects of these practices and processes may be applicable in ML research, including parti cularly: external expert evaluation of risks and appropriate responses, coordinated/responsible disclosure, trai ning in responsible research processes, disclosure coordi nation entities, safety level designations, safety level defining entities, and case-dependent response (depending on safet y levels). Reframing and renaming all of these practices and processes to focus on the release of potentially hazardous ML systems leaves us with the following components that may be n eeded in ML: •Release options : Different options for release. •Release rubric : Guidelines for when to use each type (decided by case-depen dent evaluation). •Release rubric processes : How to do case-dependent evaluation. •Release coordination : Who decides/gets access.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

22

•Release rubric processes : How to do case-dependent evaluation. •Release coordination : Who decides/gets access. •Release training : How to learn processes/norms. •Release process entities : Who manages all of this? Each of these components can be broken down further; we explo re “release options” here as an example. 6 DRAFT - JULY 2019 3.4 Release options The question of how to release research with potential for ma l-use is not a binary one: there are many different choices to make beyond simply ‘release’ or ‘don’t release’. Focusin g on this binary choice can lead the debate around openness of ML research to become very polarized. Some important dimensions we might consider when thinking a bout release strategies include: •Content: What is released Potential options include: –A fully runnable system (with varying power) –A modifiable system (with varying modifiability) –Source code (varying versions) –Training data (varying sizes) –Trained models (varying strength/fine-tunability/data-n eeds) –Paper/concept (varying detail level) –Harmful use case ideas (varying detail level)

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

23

–Paper/concept (varying detail level) –Harmful use case ideas (varying detail level) •Timing: When it is released Potential options include: –Immediate release –Timed release: Set a specific time to release components, all owing time for mitigation of any potential harms. This is common in information security. –Periodic evaluation: Don’t release immediately, but set a t ime period/intervals (e.g. every 2 months), at which point an evaluation is done to reassess the risk of rele ase given mitigation progress. –Evented release: Wait to release until some particular type of external event (e.g. someone else replicating or publicizing the same technology). –Staged release: Release systems of successively increasin g levels of power on a fixed timeline, or triggered by external events. •Distribution: Where/Who it is released to Potential options include: –Public access (with varying degrees of publicity) –Ask for access: Anyone who wants access to data or a system ask s and is approved on a case-by-case basis, potentially with specific requirements around use. –Release safety levels: People and possibly organizations c an request to be recognized as ‘safe’, after

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

24

basis, potentially with specific requirements around use. –Release safety levels: People and possibly organizations c an request to be recognized as ‘safe’, after auditing and approval they gain the ability to access all mat erial at a given safety level. –Access communities: Research groups developing their own t rusted communities through informal processes which all have access to shared repositories. Within the domain of synthetic media, it’s worth diving deep er into potential mitigations specific to products, models, and demos relevant to that space. There are a number of mechanisms researchers and companies c an use to reduce malicious use from general synthetic media systems that allow e.g. virtual impersonation: •Consent : Requiring consent by those being impersonated. •Detectability : Intentionally not trying to thwart detection. •Watermarking : Embedding context about modifications/original. •Referenceability : Centrally storing all modifications for reference. It’s important to note that none of these are perfect – they ar e part of a “defense in depth”. It is also possible to add constraints on synthesis (e.g. ensuring that only particul ar faces can be generated through the system). 7 DRAFT - JULY 2019

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

25

constraints on synthesis (e.g. ensuring that only particul ar faces can be generated through the system). 7 DRAFT - JULY 2019 3.5 Examples in practice This menu of options is valuable in theory, but it’s also wort h briefly exploring some examples in practice. One of the most notable public positions in this space comes from Go ogle: “We generally seek to share Google research to contribute to growing the wider AI ecosystem. However we do n ot make it available without first reviewing the potential risks for abuse. Although each review is content spe cific, key factors that we consider in making this judgment include: risk and scale of benefit vs downside, nature and uni queness, and mitigation options.” [18] Beyond Google, a number of labs have had to consider these iss ues. As mentioned earlier, the researchers behind Face2Face and those behind many other synthetic media syste ms have chosen not to share their code. Some researchers have released code but intentionally made it difficult to use for non-experts. Different product companies in this space are also explorin g mitigations.6Synthesia is only working with closely

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

26

Different product companies in this space are also explorin g mitigations.6Synthesia is only working with closely vetted clients. Lyrebird, which enables voice cloning, mak es it more difficult to impersonate someone without consent by requiring users to speak particular phrases instead of ju st training on arbitrary provided data. 4 Disagreements around release practices Different people and groups will have differing views on whi ch kinds of release strategies should be used when. Here we lay out some different dimensions on which people may disa gree which affect their views about release strategies for ML research. Our aim is to recognize that genuine divides exist and can lead to polarization of opinion, but that more nuanced discussion can prevent this. 4.1 Value trade-offs Some disagreements stem from fundamental views about the va lue of openness vs. caution in research. The ML community has very strong norms around openness: free sharing of data, algorithms, models, and research papers. These strong openness norms appear to be broadly mot ivated by (1) distributing the benefits of research widely by making it accessible to all of society, and (2) enab ling scientific progress by making it easier for researchers to critique and build on one another’s work.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

27

to critique and build on one another’s work. Research practices that attempt to limit mal-use by being mo re cautious about how it is released and distributed necessarily reduce some forms of openness. Some who take ope nness to be a fundamental value in research may therefore disagree with such practices on principle. Howev er, there are multiple different aspects to openness in research, and, as we’ve tried to highlight in this paper, mul tiple different approaches to being cautious about researc h release. Not all of these will necessarily be in tension with one another, and more exploration of research practices that decrease risk while protecting the most important aspe cts of openness would be valuable. 4.2 Beliefs about risks Some disagree about the relative size of different risks inv olved in ML research. On the one hand, there is the risk that advances in ML might be m isused by malicious actors in potentially catastrophic ways, which we’ve discussed. But restricting the release of ML research also creates its own risks: (1) of increasing power concentration, as a few research groups disproportio nately control how ML capabilities evolve, and (2) of creating public confusion or even panic, by creating the imp ression that advances are more threatening than they are.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

28

creating public confusion or even panic, by creating the imp ression that advances are more threatening than they are. Beliefs about the relative size of these risks can lead to two very different perspectives. Those who believe that ML advances will lead to significant harm very soon may want to ri sk such power concentration in order to safeguard democracy and public trust in the long term. By contrast, for those who think weaponization is less immediately relevant and that we can reassess risks in the future, the cos ts of restricting research may seem less palatable. While there is a genuine tension here, it is worth considerin g approaches that could address both sides of the concern (or at least address one side without exacerbating the other .) For example, some standardization of release practices, potentially managed by external entities, could help mitig ate misuse without leading to power concentration. 6For more on this, see this crowdsourced list of organization s and their actions to mitigate risk: http://bit.ly/synth- ethics. 8 DRAFT - JULY 2019 4.3 Beliefs about efficacy Another dimension of disagreement centers not around what t he risks are but how effective different practices are likely to be at reducing them. Given strong incentives or low barriers to develop a technol ogy (or achieve an insight), some suggest it is impossible

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

29

likely to be at reducing them. Given strong incentives or low barriers to develop a technol ogy (or achieve an insight), some suggest it is impossible to prevent either from leading to mal-use the long run, which could mean that restricting the release of research with potential for mal-use is futile. Others suggest that we can s ignificantly impact incentives or barriers, or that slowing down release into the world can still make a significant diffe rence, especially if this gives us time to build defenses against potential mal-use. There is also the perspective th at it is easier to build systems to defend against mal-use if more research is public, and the counterview that public inf ormation can sometimes help attackers more than defenders (‘security through obscurity’ may be unnecessary for e.g. k eeping data private but is still allegedly crucial for antispam defense). As ML researchers continue to experiment wit h release practices and explore similar challenges in other fields, we may learn about the efficacy of different appr oaches which can help inform these beliefs. 4.4 Beliefs about future needs Finally, there’s a question of whether we might eventually n eed processes for release of ML research, even if they’re not essential now.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

30

Finally, there’s a question of whether we might eventually n eed processes for release of ML research, even if they’re not essential now. For those who believe that we might develop much more advance d ML systems in the relatively near future, and that potential for harm will increase with these advances, t hen it probably makes sense to start developing careful norms and processes now regardless of current harms. For tho se who are more skeptical of the possibility of much more advanced capabilities, think that such capabilities a re unlikely to be dangerous, and/or that restricting releas e is unlikely to be effective in the future regardless, developi ng such processes now looks unnecessary. * Part of the reason for laying out various different options f or release of research is to show that this needn’t be a polarized debate: it’s not a simple choice between ‘open’ or ‘closed’ ML research. It’s worth considering whether, within our menu of options, there are approaches which can st rike a balance between the differing perspectives outlined here. 5 Recommendations We’ve laid out some considerations, tools, and options for t hinking through release of potentially harmful research in a nuanced way. But what must be done now? Here are some brief re commendations: 1.Increase understanding of the risk landscape and possible mitigation strategies:

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

31

a nuanced way. But what must be done now? Here are some brief re commendations: 1.Increase understanding of the risk landscape and possible mitigation strategies: •Develop standardized language for talking about these issues e.g. around hazards, adversa ries, mitigations and release options. •Map risks of different types of ML research in collaboration with subj ect matter experts, such as e.g. misinformation security researchers for synthetic media. Map out both immediate direct threats and potential longer-term path dependencies, in ways that addr ess researcher concerns around risk hyperbole. Develop practices for safely discussing such risks.7 •Map mitigation options , e.g. ways of reducing the harms resulting from mal-use of sy nthetic media research, and the stages/times at which they are applicable . 2.Build a community and norms around competency in understanding the impacts of ML resear ch: •Establish regular workshops to focus on release challenges. •Spread awareness of the risks of ML research to both groups who might be affecte d and who can help mitigate the risks. Proactively seek to include and learn fr om those who have been impacted. •Encourage impact evaluation both positive and negative, for research publications, pre sentations, and proposals (such as that proposed by the ACM FCA [19]).

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

32

•Encourage impact evaluation both positive and negative, for research publications, pre sentations, and proposals (such as that proposed by the ACM FCA [19]). 3.Fund institutions and systems to grow and manage research practices in ML, including poten tially: 7This type of investigation might also be referred to as e.g. t hreat models, risk analysis, or impact analysis, each of whi ch involves a different set of useful lenses. 9 DRAFT - JULY 2019 •Support expert impact evaluation of research proposals, so that the burden of this does not fal l entirely on individual researchers (who may not have the relevant exp ertise to assess hazards). This might involve e.g. identifying groups with subject matter expertis e who can do evaluations (at the request of researchers), coordinating, and potentially even paying f or review. •Prototype vetting systems to help enable shared access to potentially sensitive resea rch (as opposed to the current system where researchers attempt to validate if those requesting their models are malicious actors [20], often via error-prone ad-hoc Googling). •Develop release procedures for research already deemed to raise potential risks (manag ing all of the

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

33

actors [20], often via error-prone ad-hoc Googling). •Develop release procedures for research already deemed to raise potential risks (manag ing all of the above if needed, so that individual researchers can spend mo re time on actual research while still mitigating risks). Currently, organizations are unilaterally not publicly releasing results, so developing better procedures could actually open up research. 6 Conclusion It is clear that advances in ML have the potential to be misuse d: the main example we have discussed here is how advances in synthetic media creation may be used to sow disin formation and mistrust (but many others can be, and have been discussed [2]). We must start thinking about how to responsibly safeguard ML research. Here we focus on the role of release and publication practice s in preventing mal-use of ML research. The idea that we might sometimes restrict research release has been met with understandable concern from parts of the ML community for whom openness is an important value. Our aim here has been to decrease polarization in this debate; to emphasize that this is not a simple choice between “open” and “closed” M L research. There are a variety of options for how and when different aspects of research are released, includ ing many drawn from parallels to existing fields, and many possible processes for making these decisions.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

34

and when different aspects of research are released, includ ing many drawn from parallels to existing fields, and many possible processes for making these decisions. There will always be disagreements about the relative risks and benefits of different types of research, the effectivene ss of different mitigation strategies, and ultimately how to b alance the values of openness vs. caution. We must more deeply explore the risks and options, and develop release st rategies and processes that appropriately balance and manage the trade-offs. Ultimately, we want research to benefit humanity. We see this work as part of a maturing of the ML community, alongside crucial efforts to ensure that ML systems are fair , transparent, and accountable. As ML reshapes our lives, researchers will continue to come to terms with their new pow ers and impacts on world affairs. References [1] Jess Whittlestone, Rune Nyrup, Anna Alexandrova, Kanta Dihal, and Stephen Cave. Ethical and Societal Implications of Algorithms, Data, and Artificial Intelligence: a Roadmap for Research. London: Nuffield Foundation , 2019.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

35

2019. [2] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Pe ter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation. arXiv preprint arXiv:1802.07228 , 2018. [3] Ali Breland. The Bizarre and Terrifying Case of the “Deep fake” Video that Helped Bring an African Nation to the Brink. Mother Jones , 2019. [4] Nic Ker. Is the political aide viral sex video confession real or a Deepfake? Malay Mail , 2019. [5] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based G enerator Architecture for Generative Adversarial Networks. In Proceedings of the IEEE Conference on Computer Vision and Pa ttern Recognition , pages 4401– 4410, 2019. [6] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody Dance Now. arXiv preprint arXiv:1808.07371 , 2018.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

36

arXiv:1808.07371 , 2018. [7] Justus Thies, Michael Zollhofer, Marc Stamminger, Chri stian Theobalt, and Matthias Nießner. Face2face: Realtime face capture and reenactment of rgb videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 2387–2395, 2016. [8] RJ Skerry-Ryan, Eric Battenberg, Ying Xiao, Yuxuan Wang , Daisy Stanton, Joel Shor, Ron J Weiss, Rob Clark, and Rif A Saurous. Towards End-to-End Prosody Transfer for E xpressive Speech Synthesis with Tacotron. arXiv preprint arXiv:1803.09047 , 2018. 10 DRAFT - JULY 2019 [9] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog , 1(8), 2019. [10] Drew Harwell. Fake-porn videos are being weaponized to harass and humiliate women: ‘Everybody is a potential target’. Washington Post , December 2018.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

37

[10] Drew Harwell. Fake-porn videos are being weaponized to harass and humiliate women: ‘Everybody is a potential target’. Washington Post , December 2018. [11] BBC News. Fake voices ‘help cyber-crooks steal cash’, J uly 2019. [12] Raphael Satter. Experts: Spy used AI-generated face to connect with targets. AP News , June 2019. [13] Claire Leibowicz, Steven Adler, and Peter Eckersley. W hen Is it Appropriate to Publish High-Stakes AI Research? Partnership on AI blog post, April 2019. [14] Nick Bostrom. Information Hazards: A Typology of Poten tial Harms from Knowledge. Review of Contemporary Philosophy , 10:44–79, 2011. [15] Gregory Goth. Propaganda Bots vs. Transparency. ACM News , 2017. [16] World Health Organization. Laboratory Biosafety Manual . Geneva: World Health Organization, 2004. [17] Kyle B Enfield and Jonathon D Truwit. The Purpose, Compos ition, and Function of an Institutional Review Board: Balancing Priorities. Respiratory care , 53(10):1330–1336, 2008. [18] Google. Perspectives on Issues in AI Governance, 2019.

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729

1907.11274

38

[18] Google. Perspectives on Issues in AI Governance, 2019. [19] Brent Hecht, L Wilcox, JP Bigham, J Schöning, E Hoque, J E rnst, Y Bisk, L De Russis, L Yarosh, B Anjum, et al. It’s Time to Do Something: Mitigating the Negative Imp acts of Computing Through a Change to the Peer Review Process. ACM Future of Computing Blog , 2018. [20] Rowan Zellers. Why We Released Grover, July 2019. We are grateful to the many researchers across academia, ind ustry, and civil society who provided invaluable feedback. This is meant to be just the start of a conversation, and we are likely to update this document as we learn more. If you have thoughts or feedback, or would like to otherwise contri bute to this discussion, please reach out to the authors via email or on Twitter. 11

1907.11274

Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning

The aim of this paper is to facilitate nuanced discussion around research norms and practices to mitigate the harmful impacts of advances in machine learning (ML). We focus particularly on the use of ML to create "synthetic media" (e.g. to generate or manipulate audio, video, images, and text), and the question of what publication and release processes around such research might look like, though many of the considerations discussed will apply to ML research more broadly. We are not arguing for any specific approach on when or how research should be distributed, but instead try to lay out some useful tools, analogies, and options for thinking about these issues. We begin with some background on the idea that ML research might be misused in harmful ways, and why advances in synthetic media, in particular, are raising concerns. We then outline in more detail some of the different paths to harm from ML research, before reviewing research risk mitigation strategies in other fields and identifying components that seem most worth emulating in the ML and synthetic media research communities. Next, we outline some important dimensions of disagreement on these issues which risk polarizing conversations. Finally, we conclude with recommendations, suggesting that the machine learning community might benefit from: working with subject matter experts to increase understanding of the risk landscape and possible mitigation strategies; building a community and norms around understanding the impacts of ML research, e.g. through regular workshops at major conferences; and establishing institutions and systems to support release practices that would otherwise be onerous and error-prone.

http://arxiv.org/pdf/1907.11274

11 pages. Language fixes and tweaks for clarity

cs.CY

20190725

20190729