CMCoref: A Constraint-Based Approach for Document Coreference Resolution

The coreference resolution task aims to identify pronouns and noun phrases that represent the same entity. However, existing methods often struggle with cross-sentence coreference relations in lengthy texts. Relying on a single semantic information feature and ignoring the correlation within the task all lead to challenges in solving the problem of coreference resolution between multiple sentences. Therefore, this study proposes a constraint-based coreference resolution method CMCoref, which comprehensively optimizes the task in two aspects, mention identification and coreference prediction. For mention identification, we integrate syntactic features and semantic knowledge to enhance mention representation and introduce a constraint-based masking mechanism to improve the accuracy of mention identification. In coreference prediction, we utilize entity type features to optimize antecedent representation and exploit syntactic features along with mention type features to reinforce internal correlation within the same coreference cluster. Experimental evaluation on dataset OntoNotes 5.0 confirms the method’s superiority in mention identification and coreference prediction, offering new insights into understanding coreference resolution tasks and advancing document-level knowledge extraction.