Coreference Resolution

The most common subtype, resolving a pronoun or referring expression back to its antecedent.

• Pronominal Anaphora: Linking 'she' to 'Dr. Smith'
• Definite Noun Phrases: Resolving 'the company' to 'Acme Corp'
• Demonstratives: Binding 'that decision' to a previously described event

This is the foundational mechanism for maintaining topic continuity across sentences.

A forward-referencing structure where a pronoun precedes the noun it refers to.

• Example: 'Although he was exhausted, John continued working.'
• Requires the model to hold a placeholder entity until the explicit mention appears
• Common in literary and formal writing to create stylistic suspense
• Computationally more demanding than anaphora due to the temporary ambiguity

The process of grouping all mentions—proper nouns, pronouns, and nominal phrases—into a single coreference chain.

• A chain for 'Microsoft' might include: 'Microsoft', 'the tech giant', 'it', 'the company', 'they'
• Uses pairwise mention scoring followed by agglomerative clustering
• Modern neural models employ span-based architectures to score all possible mention spans simultaneously
• Output is a partition of the document's entity space

A benchmark requiring deep commonsense reasoning to resolve ambiguous pronouns.

• Classic example: 'The city council refused the demonstrators a permit because they feared violence.'
• 'They' refers to the council if fearing violence means fearing a riot; it refers to demonstrators if fearing police brutality
• Solving this demands world knowledge beyond syntactic patterns
• GPT-4 class models now achieve near-human performance on this task

Handling languages like Japanese, Chinese, and Korean where pronouns are frequently omitted entirely.

• The subject or object is dropped when inferable from context
• Requires the model to detect a gap in the syntactic structure and fill it with the correct entity
• Critical for multilingual conversational AI systems
• Often solved with empty category detection in syntactic parse trees

Linking mentions of the same entity across multiple documents rather than within a single text.

• Essential for knowledge base population and intelligence analysis
• Resolves 'President Biden' in one article to 'Joseph R. Biden' in another
• Uses entity linking to canonical knowledge graph IDs (e.g., Wikidata Q6279)
• Requires robust named entity disambiguation to handle shared names

The algorithmic process of disambiguating and linking textual mentions to a single, canonical entity within a knowledge base. While coreference resolution links pronouns to named entities within a document, entity resolution maps those entities to a unique identifier (e.g., a Wikidata QID).

• Within-document: Resolving 'she' to 'Ada Lovelace'
• Cross-document: Determining that 'Ada Lovelace' in document A and 'Lady Lovelace' in document B refer to the same real-world person
• Key technique: Uses entity linking algorithms that compare contextual embeddings against a canonical knowledge graph

Two directional subtypes of coreference that define how referring expressions relate to their antecedents.

• Anaphora: The referring expression points backward to a previously mentioned entity. Example: 'Marie Curie won the Nobel Prize. She was a pioneer.'
• Cataphora: The referring expression points forward to an entity not yet introduced. Example: 'When he arrived, John was tired.'
• Zero Anaphora: Common in pro-drop languages like Japanese, where the pronoun is omitted entirely and must be inferred from context

The process of maintaining a structured representation of user goals, intents, and slot values across multiple turns in a conversational AI interaction. DST depends heavily on coreference resolution to map pronouns and definite descriptions back to previously mentioned entities.

• Slot filling: 'Find a flight to Paris' → destination: Paris
• Coreference in DST: 'Book the first one' must resolve 'one' to a previously mentioned flight option
• Joint goal accuracy: The standard metric measuring whether all slot values are correctly tracked across a dialogue

A benchmark specifically designed to test pronoun disambiguation that requires real-world commonsense reasoning. Unlike statistical coreference, Winograd schemas cannot be solved through surface-level co-occurrence patterns.

• Classic example: 'The city council refused the demonstrators a permit because they feared violence.' Who feared violence?
• Requires: Understanding that city councils (not demonstrators) typically fear violence
• Significance: GPT-4 achieves ~96% accuracy, demonstrating that modern LLMs encode substantial world knowledge for pronoun resolution

The prerequisite step to coreference resolution that identifies all noun phrases, named entities, and pronouns that refer to real-world entities in a text. Without accurate mention detection, coreference chains cannot be built.

• Mention types: Named entities ('Barack Obama'), nominal phrases ('the president'), pronouns ('he')
• Singleton detection: Identifying entities mentioned only once that do not participate in coreference chains
• Nested mentions: Handling phrases like 'the president of the United States' where one mention contains another

The system architecture component responsible for storing, summarizing, and recalling dialogue history to maintain stateful context across turns. Coreference resolution is the linguistic engine that enables memory systems to track which entities are being discussed.

• Short-term memory: The current session's entity map and coreference chains
• Long-term memory: Persistent user profiles with canonical entity references
• Integration: RAG systems use coreference resolution to expand queries with resolved entity names before vector retrieval, improving recall for follow-up questions