Coreference Resolution

Coreference resolution handles two primary directional relationships: anaphora, where a pronoun refers back to a previously mentioned entity ('The court ruled... it found'), and cataphora, where a pronoun precedes its referent ('Before he testified, John reviewed the contract'). Legal texts frequently employ cataphora in introductory clauses and statutory preambles, making bidirectional resolution essential for accurate fact extraction.

The core algorithmic output is a set of coreference chains—clusters of mentions that co-refer. For example:

• Mention 1: 'Acme Corp.'
• Mention 2: 'the Defendant'
• Mention 3: 'it'
• Mention 4: 'the corporation' All four are resolved into a single entity cluster. Modern systems use span-based neural ranking to score candidate antecedents for each mention, selecting the highest-probability link to build these chains incrementally.

Legal texts are dense with Winograd schemas—ambiguous pronoun references that require world knowledge and logical reasoning to resolve. Example: 'The city council refused the demonstrators a permit because they feared violence.' Resolving 'they' requires understanding that councils, not demonstrators, typically fear violence. Domain-specific legal pre-training on case law corpora significantly improves performance on these syntactically identical but semantically distinct constructions.

In multi-document legal reasoning, entities must be linked across document boundaries. A party mentioned in a complaint, a deposition transcript, and a final judgment must be recognized as the same entity. This requires entity linking to a canonical knowledge base or legal ontology, often using embeddings to match mentions despite surface-form variations like 'Robert J. Smith,' 'Mr. Smith,' and 'the decedent.'

Beyond entity pronouns, legal coreference must resolve event references ('This ruling established...') and discourse deixis ('For the foregoing reasons...'). These abstract anaphors point to clauses, arguments, or entire sections rather than named entities. Failure to resolve them causes downstream summarization systems to lose the logical connective tissue binding a judicial opinion's reasoning chain together.

Coreference resolution systems are evaluated using MUC, B³, and CEAF scores, which measure how accurately predicted entity clusters match gold-standard annotations. The CoNLL-2012 shared task established the standard averaging of these three metrics. For legal domains, LEGAL-BERT fine-tuned on annotated court decisions achieves state-of-the-art results, with F1 scores exceeding 85% on in-domain test sets.