Binding Theory

An anaphor (reflexive or reciprocal like 'himself' or 'each other') must be bound within its local domain—typically the minimal clause containing it, its governor, and an accessible subject.

• Binding: An anaphor requires a c-commanding antecedent within the same clause
• Example: In 'John_i hurt himself_i,' the reflexive is locally bound
• Violation: 'John_i thinks [Mary_j hurt himself_i]' is ungrammatical because the antecedent is outside the local domain
• NLP Application: Principle A filters impossible antecedents for reflexives, reducing the candidate search space

A pronominal (non-reflexive pronoun like 'him' or 'her') must be free within its local domain—it cannot be bound by a c-commanding antecedent in the same minimal clause.

• Complementary distribution: Pronominals and anaphors occupy opposite binding environments
• Example: 'John_i thinks [Mary_j likes him_i]' is grammatical because 'him' is free in the embedded clause
• Violation: 'John_i likes him_i' is ungrammatical if 'him' refers to John
• NLP Application: Principle B prevents a coreference system from incorrectly linking a pronoun to a local subject

A referring expression (R-expression)—a full noun phrase like 'John' or 'the president'—must be free everywhere. It cannot be bound by any c-commanding antecedent at any level of structure.

• Global constraint: Unlike Principles A and B, this applies across all clauses
• Example: 'He_i thinks [John_i is smart]' is ungrammatical if 'He' and 'John' corefer
• Cataphora restriction: A pronoun cannot c-command a coreferring R-expression that follows it
• NLP Application: Principle C blocks cataphoric coreference where a pronoun precedes and c-commands its antecedent

C-command (constituent command) defines the structural relationship that determines binding possibilities. Node A c-commands node B if A does not dominate B and the first branching node dominating A also dominates B.

• Syntactic asymmetry: C-command encodes hierarchical prominence, not linear order
• Example: In a standard tree, a subject c-commands objects within the same clause
• Binding condition: Binding requires both c-command and co-indexation
• NLP Application: C-command relationships, derived from dependency parses, provide hard structural features for neural coreference models to prune impossible antecedents

The governing category—the minimal domain for Principles A and B—is defined as the minimal clause containing the governed element, its governor (a lexical head like a verb or preposition), and an accessible SUBJECT (a structural subject or agreement features).

• Governor: The lexical head that assigns case or theta-role to the dependent element
• Accessible SUBJECT: Can be an overt subject, a tense/agreement marker, or the anaphor itself in certain constructions
• Cross-linguistic variation: Languages differ in what counts as an accessible SUBJECT, affecting binding domain boundaries
• NLP Application: Accurate governing category identification requires syntactic parsing and informs language-specific coreference constraints

Modern neural coreference systems incorporate Binding Theory constraints as hard filters or soft features to improve precision by eliminating linguistically impossible antecedents.

• Hard constraint filtering: Remove candidate antecedents that violate Principles A, B, or C before scoring
• Soft feature integration: Encode binding constraint violations as negative features in the mention-ranking scoring function
• Complementary to learned representations: Binding constraints provide structural knowledge that may not be reliably learned from limited training data
• Example: In 'John told Bill about himself,' Principle A restricts 'himself' to 'Bill' (the local subject), not 'John'

A linguistic expression whose interpretation depends on a preceding expression (the antecedent). Typically, a pronoun referring back to a previously mentioned noun phrase.

• Example: In 'Sarah lost her keys,' 'her' is an anaphor referring to 'Sarah.'
• Role in NLP: The primary target of pronominal resolution systems.
• Contrast: Differs from cataphora, where the pronoun precedes its referent.

The complete ordered set of all mentions within a discourse that refer to a single real-world entity. It forms a linked sequence from the first mention to the last.

• Structure: [Mention 1] -> [Mention 2] -> [Mention 3]
• Example: 'Alice said she would call. The manager was busy.'
• Importance: The gold-standard output for evaluating end-to-end coreference resolution systems like the e2e-coref model.

A discourse coherence theory that models local focus to constrain pronoun interpretation. It tracks a ranked set of forward-looking centers (potential future topics) and a single backward-looking center (the current topic).

• Constraint: The preferred antecedent for a pronoun is the highest-ranked forward-looking center from the previous utterance.
• Application: Used as a feature in rule-based sieve architectures and salience models to improve pronominal resolution accuracy.

A neural coreference architecture that scores all candidate antecedents for a given mention and selects the highest-ranked one. This replaces independent pairwise decisions with a global ranking.

• Scoring: Uses biaffine attention to compute pairwise scores between mention and antecedent representations.
• Refinement: Supports higher-order inference, where span representations are iteratively updated based on predicted antecedents to enable transitive reasoning across chains.
• Efficiency: Relies on antecedent pruning and span pruning to restrict the candidate search space.

A pronoun disambiguation challenge requiring deep world knowledge and commonsense reasoning. Two sentences differ by a single word that flips the pronoun's antecedent.

• Example: 'The city council refused the demonstrators a permit because they feared violence' vs. '...because they advocated violence.'
• Significance: Designed to be statistically trivial for language models but requires genuine understanding, making it a robust test for evaluating commonsense reasoning in coreference systems.

A non-identity anaphoric relationship where a definite noun phrase refers to an entity inferentially linked to a previously introduced discourse referent, rather than directly coreferring with it.

• Example: 'I bought a car. The engine is loud.' — 'the engine' is not coreferent with 'a car' but is a part-of relation.
• Challenge: Requires world knowledge beyond syntactic constraints, making it significantly harder to resolve than standard identity coreference.
• Contrast: Differs from split antecedents, where a plural pronoun directly refers to multiple distinct entities.