| # 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} | |
| } | |
| ``` |