Train and evaluate a transformers based model for the stance detection task between the claims and the related documents.
The covidex engine should highlight the most relevant sentences in each related document, so we shouldn't need too do that extraction ourselves. Different hyperparams and data manipulations should be tested, but as a baseline, we should try:
roberta-large with hyperparams used in the roberta paper
text_a is the claim, and text_b will be the highlighted sentences
Another thing that should be looked into is already pretrained models for stance detection that might be out there. Possibly using slightly different architectures etc ...
Train and evaluate a transformers based model for the stance detection task between the claims and the related documents.
The covidex engine should highlight the most relevant sentences in each related document, so we shouldn't need too do that extraction ourselves. Different hyperparams and data manipulations should be tested, but as a baseline, we should try:
Another thing that should be looked into is already pretrained models for stance detection that might be out there. Possibly using slightly different architectures etc ...