Coreference Resolution: Linking Pronouns to Their Antecedents

Passage

—

Scoring weights

Semantic ↔ Recency balance 0.55

Recency decay rate τ 0.35

Enforce gender/number agreement

Resolving

—

Passage accuracy

—

Top-1 vs top-2 margin

—

Pronoun

—

Candidates scored

—

How it works

For a pronoun mention p and every earlier mention c in the passage, a mention-pair coreference model scores how likely c is the antecedent of p:

sim(p,c)  = cos(embed(p), embed(c))        ∈ [-1, 1]
rec(p,c)  = exp(-τ · (wordPos(p) - wordPos(c)))
score(p,c)= w·((sim+1)/2) + (1-w)·rec
if agreement(p,c) fails:
   score ← -∞   (hard filter, if enforced)
   score ← score · 0.35   (soft penalty, otherwise)
antecedent(p) = argmax_c score(p,c)

Semantic ↔ Recency balance w — trades off two real coreference features: embedding similarity between the pronoun's context and a candidate entity (does the meaning fit?) versus how recently that candidate was mentioned (nearby mentions are cheaper to track and more often correct).
Recency decay τ — how fast the recency signal falls off with word distance; a high τ makes the model almost only trust the immediately preceding mention.
Agreement — real coreference resolvers use gender/number/animacy as a hard linguistic constraint ("they" can't be a singular person) before any learned score is even computed. Turning it off shows how often semantics and recency alone still get it right — and where they don't.
The dog/cat/tree passage keeps agreement useless on purpose (every candidate is grammatically neuter "it"), so only semantic similarity and recency can disambiguate — exactly the hard case real coreference systems (e.g. end-to-end neural coreference, spaCy's neuralcoref, or transformer span-based resolvers) are built to handle.

Arcs in the scene fan out from the pronoun currently being resolved back to every legal candidate; arc brightness and thickness are proportional to that candidate's score, and the winner is highlighted. Box color shows the model's predicted cluster for every mention so you can see full-passage accuracy at a glance.

Entity A Entity B Entity C Unresolved candidate

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.