Salience Model

Entities mentioned more recently in the discourse receive higher salience scores. This is based on the cognitive principle that short-term memory retains recently activated concepts. In practice, a linear decay function reduces an entity's prominence as new sentences are processed. For example, in "John went to the store. He bought milk," John is highly salient when "He" is encountered because he was mentioned in the immediately preceding clause. Recency is often implemented as a distance metric counting the number of sentences or clauses since the entity's last mention.

Entities occupying prominent syntactic positions are assigned higher salience. The typical hierarchy is:

• Subject > Object > Oblique/Other
• Entities in subject position are more likely to be the backward-looking center and serve as antecedents for subsequent pronouns.
• In "Mary gave the book to Susan. She smiled," Mary (subject) is weighted higher than Susan (oblique object), making Mary the preferred antecedent for "She." This reflects the linguistic universal that subjects encode topic continuity.

Entities that are mentioned repeatedly accumulate higher salience through frequency reinforcement. Each additional mention of an entity increases its cumulative prominence score, making it more resistant to decay. This models the discourse phenomenon of topical persistence: the main topic of a passage is referenced multiple times and remains the preferred antecedent even when other entities intervene. For instance, a central character in a narrative will maintain high salience across paragraphs due to repeated mentions, while incidental entities fade quickly.

Salience models operationalize Centering Theory by maintaining an ordered list of forward-looking centers (Cf) ranked by prominence. The highest-ranked entity becomes the preferred backward-looking center (Cb) of the next utterance. A key constraint is that if any pronoun appears in an utterance, it must refer to the Cb. This creates a predictive framework: the salience ranking directly determines which entity a pronoun will resolve to. Transitions between utterances—continue, retain, smooth-shift, rough-shift—are classified based on how the Cb and Cf rankings change.

Modern neural salience models combine multiple features into a learned scoring function rather than relying on hand-crafted heuristics. Typical input features include:

• Binary recency flags (mentioned in last N sentences)
• Syntactic depth in the parse tree
• Mention type (proper name > definite description > pronoun)
• Parallelism with the anaphor's grammatical role These features are fed into a feedforward network that outputs a scalar salience score, which is then used to rank candidate antecedents in a mention-ranking architecture.

Advanced salience models incorporate discourse segmentation to handle paragraph and section boundaries. Entities are assigned salience within a local discourse segment, and a separate global salience tracks prominence across segments. When a discourse boundary is crossed—such as a new section or a temporal shift—local salience scores are partially reset or decayed at an accelerated rate. This prevents entities from earlier, unrelated passages from incorrectly competing with newly introduced referents, improving resolution accuracy in long documents.

A discourse coherence theory that models local focus by tracking a ranked set of forward-looking centers (Cf) and a single backward-looking center (Cb). The theory posits that the most highly ranked entity in the current utterance is the preferred antecedent for a pronoun in the next utterance. Salience is determined by grammatical role hierarchy: Subject > Object > Other. Transitions between utterances (Continue, Retain, Shift) are evaluated based on whether the Cb remains the same, directly constraining pronoun interpretation.

A neural coreference architecture that operationalizes salience by scoring all candidate antecedents for a given mention and selecting the highest-ranked one. Unlike pairwise models, this approach naturally captures the relative prominence of entities in the discourse. Key features include:

• Biaffine attention to compute pairwise scores between mention and antecedent representations
• Span representations encoding the syntactic and semantic properties of each mention
• Higher-order inference to iteratively refine entity prominence based on predicted coreference links

A syntactic theory governing the distribution of anaphors (reflexives, reciprocals), pronominals, and referring expressions. It provides structural constraints that interact with salience models by defining domains where certain coreference relationships are obligatory or prohibited. The three principles are:

• Principle A: An anaphor must be bound within its local domain
• Principle B: A pronominal must be free within its local domain
• Principle C: A referring expression must be free everywhere These constraints serve as hard filters in coreference systems, overriding salience-based preferences when syntactic rules dictate.

A rule-based method for identifying the syntactic head word of a mention span, which is critical for computing salience features. The head word carries the core semantic content and determines agreement features (number, gender, animacy) used in antecedent matching. Common heuristics traverse dependency parse trees to locate the head, prioritizing:

• Nouns over adjectives in noun phrases
• Proper nouns over common nouns
• Rightmost content words in complex spans Accurate head identification directly impacts the quality of salience scoring and pronoun resolution.

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. Unlike standard salience-driven resolution, bridging relies on world knowledge and lexical semantics. Example: 'We entered a restaurant. The waiter took our order.' The waiter is not coreferent with the restaurant but is inferentially accessible through the restaurant frame. Salience models must distinguish between identity coreference and bridging to avoid false links.

A linguistic phenomenon where a demonstrative pronoun (this, that) refers to an abstract entity, event, or proposition described in a preceding clause or sentence rather than a concrete noun phrase. Example: 'The company missed its earnings target. This caused the stock to plummet.' The pronoun 'this' refers to the entire event of missing earnings. Salience models must extend beyond entity tracking to include propositional salience, recognizing that events and abstract concepts can serve as antecedents for deictic references.