Antecedent Pruning

Restricts the search space by enforcing a maximum sentence distance between a mention and its candidate antecedents.

• Window Size: Typically limits candidates to the current and preceding 2-3 sentences.
• Rationale: Entities in natural discourse tend to be referenced locally; distant antecedents are statistically rare.
• Implementation: A sliding window over the document graph that discards spans outside the threshold before the scoring phase.

Applies linguistic rules to eliminate grammatically impossible antecedents, preventing the model from wasting compute on invalid pairs.

• Binding Theory: Filters candidates that violate Principle B (a pronoun cannot be bound in its governing category) or Principle C (a referring expression cannot be bound).
• C-Command: Removes antecedents that do not c-command the anaphor in the parse tree.
• Head Agreement: Prunes candidates where the syntactic head does not match the expected constituent type.

Eliminates candidates that fail to match the mention on morphosyntactic features, a high-recall filter applied before neural scoring.

• Number Agreement: A singular pronoun cannot corefer with a plural noun phrase.
• Gender Agreement: In gendered languages, filters antecedents that conflict with the pronoun's grammatical gender.
• Animacy: Distinguishes between animate and inanimate entities to resolve pronouns like 'it' vs. 'he/she'.
• Person: Ensures first-person pronouns are only matched with first-person antecedents.

Prioritizes or eliminates candidates based on their syntactic type, leveraging the observation that certain mention types are more likely to serve as antecedents.

• Proper Noun Priority: Named entities are often discourse-prominent and serve as antecedents for subsequent pronouns.
• Indefinite Filtering: Indefinite noun phrases (e.g., 'a company') typically introduce new entities rather than referring back.
• Appositive Collapse: Treats appositive constructions as a single mention to avoid redundant candidate generation.

A learned preprocessing step where a lightweight scorer assigns a mention likelihood to every possible span, retaining only the top-k candidates.

• Top-k Selection: Typically keeps 0.4 to 0.5 spans per token on average, reducing the candidate pool by over 90%.
• Joint Pruning: The same mention score used for antecedent pruning also defines the final mention detection threshold.
• Efficiency Gain: Reduces the complexity of the pairwise scoring matrix from O(n²) to O(n × k), where k is the pruned span count.

Rather than scoring all valid candidates, the model only evaluates antecedents in order of discourse prominence, stopping early when a high-confidence link is found.

• Recency Ranking: Sorts candidates by proximity, assuming the most recent compatible antecedent is the most likely.
• Grammatical Role Weighting: Subjects and objects are ranked higher than oblique arguments based on Centering Theory.
• Early Termination: The first candidate exceeding a confidence threshold is selected, bypassing the need to score the entire filtered set.

The dominant neural architecture where antecedent pruning is most critical. For each detected mention, the model scores all candidate antecedents in a truncated window and selects the highest-ranked one.

• Scoring Function: Uses biaffine attention or a feedforward network over concatenated span representations
• Pruning Dependency: Without aggressive pruning, the pairwise scoring matrix becomes computationally intractable for long documents
• End-to-End Training: Jointly optimizes mention detection and antecedent scoring, making efficient candidate generation essential

A preprocessing step that drastically reduces the number of candidate mentions before antecedent search begins. Only spans with a high mention score are retained.

• Mention Score: A learned scalar indicating how likely a span is to be an entity mention
• Typical Threshold: Keep only the top λT spans by mention score, where λ is a hyperparameter (often 0.4)
• Impact: Reduces candidate pool from O(n²) possible spans to a manageable few hundred, directly limiting antecedent candidates

A rule-based method for identifying the syntactic head of a mention span, used to enforce agreement constraints during antecedent pruning.

• Head Extraction: For a span like 'the large language model', the head is 'model'
• Agreement Filtering: Prunes antecedents that disagree in number (singular/plural) or gender with the current mention
• Feature Engineering: The head word provides critical features for pairwise coreference scoring in traditional mention-pair models

An iterative refinement technique where span representations are updated based on their predicted antecedents, enabling transitive reasoning across coreference chains.

• Iterative Refinement: After the first pass, each span's representation is weighted by the representations of its top antecedents
• Pruning Interaction: The antecedent distribution from the first pass informs which candidates are considered in subsequent refinement passes
• Transitive Closure: Allows the model to implicitly link mentions that share a common antecedent, improving chain consistency

A discourse model that assigns a real-valued prominence score to each entity, guiding which antecedents are most likely to be referenced by upcoming pronouns.

• Recency Weighting: Entities mentioned more recently receive higher salience scores
• Grammatical Role: Entities in subject position are more salient than those in object position
• Pruning Heuristic: Antecedent candidates can be filtered to retain only the most salient entities, dramatically reducing the search space for pronominal resolution

The standard evaluation dataset for coreference resolution derived from OntoNotes 5.0, used to measure the impact of pruning strategies on end-to-end performance.

• Metrics: MUC, B³, and CEAFφ4 evaluate different aspects of coreference chain quality
• Average F1: State-of-the-art models achieve approximately 80%+ average F1 on the English test set
• Pruning Trade-off: Evaluates the balance between computational efficiency and recall—aggressive pruning may drop recall on long-distance coreference chains