Sentence-aware Contrastive Learning for Open-Domain Passage Retrieval

Training dense passage representations via 001 contrastive learning has been shown effective 002 for Open-Domain Passage Retrieval (ODPR). 003 Existing studies focus on further optimizing 004 by improving negative sampling strategy or ex-005 tra pretraining. However, these studies keep 006 unknown in capturing passage with internal 007 representation conflicts from improper model-008 ing granularity. This work thus presents a re-009 fined model on the basis of a smaller granular-010 ity, contextual sentences, to alleviate the con-011 cerned conflicts. In detail, we introduce an 012 in-passage negative sampling strategy to en-013 courage a diverse generation of sentence rep-014 resentations within the same passage. Experi-015 ments on three benchmark datasets verify the 016 efficacy of our method, especially on datasets 017 where conflicts are severe. Extensive experi-018 ments further present good transferability of 019 our method across datasets.