###Contribution
- Propose that current relation extraction models are lack of cross-sentence and cross-mention consideration which occured frequently.
- Point out old models are inefficient by independently extracting features and score each pair.
- Introducing BRAN which synthesize convolutions and self-attention(modified on transformer). This model can score all pairs of mentions in parallel using a bi-affine operator and have some sort of way to combine them.
- Their model are jointly training relation extranction and NER. It seems to bring some recall increase.
Details of the model
- BRAN will first tokenize the text by using subword to combine words. I guest they did this because there are many “new” words in the biological articles.
- They use multi-instance learing which will consider all pairs together by using attention
- Joint training NER will have some increase of recall score
近期评论