Coreference Resolution

The most common sub-task, resolving anaphors (referring expressions) to their antecedents (prior mentions).

• Pronominal Anaphora: Resolving 'she' or 'it' to a specific noun phrase.
• Definite Noun Phrases: Linking 'the company' back to a previously introduced entity.
• Zero Anaphora: Identifying omitted subjects in pro-drop languages like Japanese or Spanish.

This process is critical for extracting coherent predicate-argument structures across sentence boundaries.

A forward-referencing phenomenon where a pronoun appears before its antecedent in the text.

• Example: 'Although he was tired, John kept working.'
• Requires bidirectional context modeling rather than simple left-to-right processing.
• Common in literary and formal writing styles.
• Poses challenges for incremental, real-time NLP systems that process text sequentially.

Modern transformer architectures handle this naturally through self-attention mechanisms.

The output of resolution is a coreference chain—a cluster of all mentions referring to a single entity.

• Example: 'Apple Inc. announced its new chip. The tech giant expects it to ship next month.'
• All four bolded mentions form one chain pointing to the same real-world organization.
• Chains are essential for building entity-centric knowledge graphs from unstructured text.
• Evaluation uses metrics like MUC, B³, and CEAF to measure cluster alignment.

Coreference extends beyond entities to abstract objects and events.

• Entity Coreference: Linking mentions of people, organizations, locations, and physical objects.
• Event Coreference: Identifying when multiple descriptions refer to the same occurrence.
  • Example: 'The merger closed on Tuesday. This transaction will reshape the industry.'
• Requires understanding of temporal relations and causal structures.
• The OntoNotes corpus annotates both entity and event coreference across multiple genres.

Entities mentioned only once in a document are singletons—they have no coreferent expressions.

• Critical distinction: a singleton is not the same as an unresolved mention.
• Many systems historically ignored singletons, but they carry important discourse weight.
• Modern end-to-end models trained on CoNLL-2012 data learn to identify all mentions, then cluster them.
• Proper singleton handling improves downstream tasks like relation extraction and knowledge base population.

A benchmark testing pronoun disambiguation requiring world knowledge and commonsense reasoning.

• Example: 'The city council refused the demonstrators a permit because they feared violence.'
• Who feared violence? Requires understanding that councils, not demonstrators, typically fear violence in this context.
• Designed as an alternative to the Turing Test, specifically targeting coreference.
• Modern large language models achieve near-human performance on this task through extensive pre-training.

A distinct phenomenon where an expression refers not to a concrete entity but to a proposition, event, or segment of discourse itself. This is often excluded from standard coreference tasks but critical for full discourse understanding.

• Example: The merger was announced on Monday. This surprised analysts.
• This refers to the entire preceding event, not a noun phrase
• Requires modeling of abstract anaphora beyond entity-level resolution
• Handled in frameworks like Abstract Meaning Representation (AMR) and discourse parsing