Span extraction and contrastive learning for coreference resolution

Coreference resolution is a primary task in natural language processing (NLP), designed to automatically identifyand classify noun phrases or pronouns that represent thesame mention. Most of the recent coreference resolution models extraction or cluster mentions. However, these models rely heavily on the word-level representation to find coreference links between words and spans. In this paper, we propose a multi-task learning approach based on the transformers, which integrates three tasks including span extraction, contrastive learning and text classification. Our model pays attention to both word and sentence representation without requiring for the complex annotations and the manual rules. Experiments demonstrate significant performance over previous methods, with an accurate 77.63% in the dev and an accurate 76.90% in the test of the CLUE-WSC.