File size: 2,358 Bytes
238b13e
 
 
 
 
 
 
 
 
 
 
 
751002f
238b13e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
# CoNTACT

### Model description

<u>Co</u>ntextual <u>N</u>eural <u>T</u>ransformer <u>A</u>dapted to <u>C</u>OVID-19 <u>T</u>weets or **CoNTACT** is a Dutch RobBERT model (```pdelobelle/robbert-v2-dutch-base```) adapted to the domain of COVID-19 tweets. The model was developed at [CLiPS](https://www.uantwerpen.be/en/research-groups/clips/) by Jens Lemmens, Jens Van Nooten, Tim Kreutz and Walter Daelemans. A full description of the model, the data that was used and the experiments that were conducted can be found in this ArXiv preprint: https://arxiv.org/abs/2203.07362

### Intended use

The model was developed with the intention of achieving high results on NLP tasks involving Dutch social media messages related to COVID-19.

### How to use

CoNTACT should be fine-tuned on a downstream task. This can be achieved by referring to ```clips/contact``` in the ```--model\_name\_or\_path``` argument in Huggingface/Transformers' example scripts, or by loading CoNTACT (as shown below) and fine-tuning it using your own code:

 ```
 from transformers import AutoModel, AutoTokenizer
 
 model = AutoModel.from_pretrained('clips/contact')
 tokenizer = AutoTokenizer.from_pretrained('clips/contact')
 
 ...
 

 ```

### Training data

CoNTACT was trained on 2.8M Dutch tweets related to COVID-19 that were posted in 2021. 

### Training Procedure

The model's pre-training phase was extended by performing Masked Language Modeling (MLM) on the training data described above. This was done for 4 epochs, using the largest possible batch size that fit working memory (32). 

### Evaluation

The model was evaluated on two tasks using data from two social media platforms: Twitter and Facebook. Task 1 involved the binary classification of COVID-19 vaccine stance (hesitant vs. not hesitant), whereas task 2 consisted of the mulilabel, multiclass classification of arguments for vaccine hesitancy. CoNTACT outperformed out-of-the-box RobBERT in virtually all our experiments, and with statistical significance in most cases.

### How to cite

```
@misc{lemmens2022contact,
    title={CoNTACT: A Dutch COVID-19 Adapted BERT for Vaccine Hesitancy and Argumentation Detection},
    author={Jens Lemmens and Jens Van Nooten and Tim Kreutz and Walter Daelemans},
    year={2022},
    eprint={2203.07362},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}
```