I would like to try classification involving an article (multiple sentences) and a claim (single sentence) to classify if the article disagrees or agrees with the claim.
I am essentially trying to replicate the project in https://arxiv.org/pdf/1911.11951.pdf (for the classification of whether a claim agrees or disagrees with a small article) where
“the dataset is pre-processed by initializing each example with a start token to signify
the beginning of a sequence, followed by the claim, two separator tokens, the article and an additional separator token”.
I was wondering what is point for two separator tokens between the claim and article is. When following the huggingface documentation it seems I just need a single [SEP] for classification involving a combination of 2 components (like article plus claim) and then theToken Type IDs.
Could it be perhaps that two separator tokens were used because Token Type IDs were not used or are Token Type IDs always required when classifying a combination of 2 components (like article plus claim)?