Acl2022 | DCSR: a sentence aware contrastive learning method for open domain paragraph retrieval

This paper introduces an article ACL2022 The paper DCSR, Sentence aware contrastive learning for open domain paragraph retrieval , By introducing 「 Sampling strategy for negative examples in paragraphs 」, Make the same paragraph generate 「 Different sentence representations 」, Construct a context sentence representation model based on smaller granularity , So as to solve the problem of , The conflict problem between the same paragraph vector and multiple problem vectors with large semantic differences .

Paper title ：

Sentence-aware Contrastive Learning for Open-Domain Passage Retrieval

Address of thesis ：

https://aclanthology.org/2022.acl-long.76.pdf

code：

https://github.com/chengzhipanpan/DCSR

The structure of the model is shown in the figure above , Two pre training models are used to solve the problem query And paragraph content passage Encoding , Insert a special... Before the sentence in each paragraph <sent> Mark . also ,「 In order to preserve the information of the text context 」, Input the paragraph content as a whole into the encoder for encoding ,<sent> Take the vector corresponding to the mark as the sentence vector .