Rule-Based Sieve

Unlike neural models that learn from data, a rule-based sieve applies hand-crafted linguistic rules that are 100% predictable and auditable. Each rule is designed to be high-precision, meaning when it fires, it is almost certainly correct. This eliminates the black-box uncertainty of learned systems, making it ideal for applications where explainability and deterministic behavior are non-negotiable, such as legal document analysis or medical record processing.

The architecture operates as a sequence of precision-ordered sieves applied from highest to lowest precision. The cascade typically includes:

• Speaker Identification: Resolves first and second person pronouns based on dialogue turns.
• Exact String Matching: Links identical mention strings (e.g., 'the company' to 'the company').
• Relaxed String Matching: Links mentions after removing determiners and modifiers.
• Precise Constructs: Applies rules for appositives, predicate nominatives, and acronyms.
• Pronoun Resolution: Uses binding theory, gender, and number agreement as the final, most aggressive sieve. Each sieve only operates on mentions not yet resolved by earlier, more reliable sieves.

A critical component of rule-based sieves is the head-finding heuristic, which identifies the syntactic head word of a mention span. For example, in 'the board of directors', the head is 'board'. This enables:

• Head matching: Comparing only the head words of two mentions for coreference.
• Feature extraction: Using the head's part-of-speech, number, and gender for agreement checks.
• Antecedent pruning: Filtering candidate antecedents whose heads conflict in number or gender. The heuristic relies on deterministic syntactic rules rather than statistical parsers, maintaining the sieve's overall determinism.

When multiple candidate antecedents satisfy a sieve's constraints, the system uses a salience model to rank them. Salience is computed deterministically based on:

• Recency: Mentions closer in the discourse are more salient.
• Grammatical role: Subjects are more salient than objects.
• Mention frequency: Entities mentioned more often are more prominent. This mirrors Centering Theory, where the most salient entity in the preceding utterance is the preferred antecedent for a pronoun. The salience ranking ensures consistent, linguistically motivated resolution when rules alone are ambiguous.

Rule-based sieves incorporate syntactic binding constraints from Chomsky's Binding Theory to filter impossible pronoun-antecedent relationships:

• Principle A: A reflexive pronoun (e.g., 'himself') must be bound within its local clause.
• Principle B: A non-reflexive pronoun (e.g., 'him') must be free within its local clause.
• Principle C: A referring expression (e.g., 'John') must be free everywhere. These constraints are implemented as hard filters that prevent a sieve from linking a pronoun to a syntactically impossible antecedent, dramatically reducing false positives in pronoun resolution.

The sieve architecture employs a multi-pass strategy where each pass resolves a specific type of coreference with increasing aggression. The passes are ordered by precision-recall tradeoff:

• Pass 1 (Speaker ID): Resolves 'I', 'you', 'we' using dialogue metadata — near-perfect precision.
• Pass 2 (String Match): Links identical or near-identical noun phrases — very high precision.
• Pass 3 (Construct Rules): Handles appositives ('John, the CEO') and acronyms ('FBI' → 'Federal Bureau of Investigation').
• Pass 4 (Pronoun Sieve): The final, most aggressive pass that resolves remaining pronouns using all available agreement and salience features. This cascading design ensures that easy cases are resolved first, leaving only the most ambiguous mentions for the most aggressive rules.