Entity A Entity B Entity C Unresolved candidate

Coreference Resolution: Linking Pronouns to Their Antecedents (2D)

Every time a language model reads "Maria met John... she smiled at him," it has to solve a small but critical puzzle: which earlier noun does each pronoun actually point to? This 2D companion renders that scoring process as a flat timeline of mention boxes: a small lexicon supplies each entity's gender and number, a simplified rule-based parser tags every mention as a grammatical subject or object from the raw token stream, and a live weighted scorer combines a hard gender/number filter with a recency-decay term and a role-salience term to rank every earlier candidate for each pronoun. Three preset passages isolate different failure modes — clean gender cues, an all-neuter "it" chain that only role and recency can untangle, and a plural "they" pair where the recency-vs-role tension is most visible — with adjustable sliders and a live per-candidate score breakdown showing exactly why the model picked the antecedent it did.