Coreference Resolution

The most common form of coreference, resolving pronouns (he, she, it, they) back to their explicit antecedents. For example, in 'Alice went to the store. She bought milk,' the system links 'She' to 'Alice.' Modern systems use span-based neural architectures that score all possible mention spans and their coreference links simultaneously, rather than processing text sequentially.

The less common forward-referencing pattern where a pronoun appears before its antecedent. Example: 'Although he was tired, John kept working.' This requires the model to hold unresolved references in memory until the explicit entity appears later in the text. Cataphora is particularly challenging for left-to-right language models that process text sequentially without lookahead capabilities.

The process of grouping all mentions—pronouns, proper names, and nominal phrases—that refer to the same real-world entity into a single coreference chain. For instance, 'Barack Obama,' 'the 44th president,' 'he,' and 'Obama' all cluster into one entity. Modern systems use agglomerative clustering over pairwise mention scores to build these chains, optimizing for global coherence rather than local decisions.

A specialized benchmark requiring world knowledge and commonsense reasoning to resolve ambiguous pronouns. Classic example: 'The city council refused the demonstrators a permit because they feared violence.' Who feared violence? The council or the demonstrators? Solving these requires understanding semantic roles, causality, and typical behavioral patterns—capabilities that test the limits of purely statistical approaches.

Identifying implicit, unpronounced references common in pro-drop languages like Japanese, Chinese, and Spanish. In 'John entered the room. [Ø] Sat down,' the zero pronoun must be resolved to 'John.' This requires the model to detect missing arguments in the syntactic structure and infer the intended referent from discourse context, adding complexity beyond explicit mention detection.

Linking mentions of the same entity across multiple documents or data sources. For example, identifying that 'IBM' in a financial report, 'International Business Machines' in a news article, and 'Big Blue' in a blog post all refer to the same organization. This is critical for knowledge base population and entity resolution pipelines that aggregate information from heterogeneous sources.

The NLP task of mapping textual mentions to their unique, unambiguous entries in a target knowledge base like Wikidata or DBpedia. While coreference resolution clusters mentions of the same entity within a document, entity linking connects those clusters to a canonical, machine-readable identifier (e.g., a Q-Node). This step is critical for transforming unstructured text into structured, queryable knowledge graph triples.

The specific sub-task of resolving which distinct real-world entity an ambiguous name refers to. For example, determining if 'Paris' refers to the capital of France or the mythological figure. Coreference resolution feeds into this by providing all the contextual mentions (e.g., 'she,' 'the city,' 'its landmarks') that disambiguation algorithms use to select the correct candidate from a knowledge base.

A core sub-task of coreference resolution focused specifically on identifying the antecedent of an anaphor—a word that refers back to a previously mentioned entity. This typically involves resolving pronouns:

• Pronominal Anaphora: Resolving 'he,' 'she,' 'it,' 'they' to their noun phrases.
• Definite Noun Phrases: Resolving 'the company' or 'the algorithm' to a specific earlier mention. Modern systems use span-based neural models to score all possible antecedent-mention pairs.

The less common but equally important resolution of a reference that points forward in the text to an entity introduced later. For example: 'Although it was designed for speed, the new processor consumed minimal power.' Here, 'it' refers cataphorically to 'the new processor.' Robust coreference systems must handle both anaphoric and cataphoric references to build complete entity chains.

An automated software workflow that ingests unstructured text and outputs structured, disambiguated entity records. Coreference resolution is a critical mid-pipeline component that operates after named entity recognition (NER) and before entity linking. By clustering all mentions of an entity, the pipeline can aggregate attributes and relationships from across a document to build a richer, more accurate knowledge graph node.

A computational method that assigns a numerical score to each entity in a document to quantify its contextual importance. Coreference resolution directly impacts salience by consolidating all mentions of an entity into a single cluster. An entity's salience score often correlates with the frequency and syntactic prominence of its coreference chain, helping downstream systems identify the document's primary subjects.