Coreference Resolution

Why Does This Exist?

Read this: "Mira adopted a stray pup. She named it Biscuit." You never wonder who She is or what it means. A pipeline without coreference does. It sees separate strings where you see two entities, so question answering answers "Who named Biscuit?" with nothing, and counts of people and pets come out doubled.

That is the concrete breakage. Pronouns carry no meaning on their own. Without a step that clusters mentions into chains, every downstream counter miscounts. A sentiment system credits "it" to nobody. An extractor emits two people named Mira and She.

This page covers the clustering step itself: finding mentions, scoring pairs, and linking chains. Finding the spans in the first place belongs to named entity recognition. Turning chains into rows belongs to information extraction. The one prerequisite is small: a mention is any span that could refer to an entity, usually a proper noun, a noun phrase, or a pronoun. Everything else on this page builds on that.

Think of It Like This

How It Actually Works

Modern systems do three things in order: detect mentions, score pairs, then cluster.

Mention detection

Every span up to some width (often 10 tokens) gets a mention score $m(i)$ from an encoder. Spans below a cutoff are dropped. So "Mira", "a stray pup", "She", "it", and "Biscuit" survive, while "adopted a" does not. This keeps the pair step quadratic in mentions, not in all spans.

Pairwise scoring

For a later mention $j$ and each earlier candidate $i$, the model computes an antecedent score $s(i, j)$. Here $h_i$ and $h_j$ are the span vectors from a transformer, $\circ$ is elementwise product, $\phi(i, j)$ holds distance and agreement features, and FFNN is a small feedforward net:

In plain words: $h_i$ says what the earlier span means in context, $h_j$ says what the later span means, their product says how similar they are, and $\phi$ says whether linking them is plausible (close together, same number and gender). Each mention also gets a no-link score $s(\epsilon, j)$ for starting a new entity. The model links $j$ to the best scorer among earlier mentions and $\epsilon$.

Worked example with real numbers

Take "Mira adopted a stray pup. She named it Biscuit." Focus on $j =$ "She" with candidate $i =$ "Mira", 5 tokens apart, both singular and feminine.

Suppose the trained weights produce $s = 2.0 \times \text{sim} - 0.1 \times \text{dist} + 0.5 \times \text{agree}$, with cosine similarity $0.82$, distance $5$, agreement $1$:

Against "a stray pup" the similarity is $0.05$, distance $3$, agreement $0$:

Softmax over $\{1.64, -0.20, 0.0\}$ for the no-link option gives exponentials $5.16$, $0.82$, $1.00$, summing to $6.98$. So $P(\text{Mira}) = 5.16 / 6.98 \approx 0.74$. "She" joins Mira with probability $0.74$, which is exactly the link the diagram draws. The same arithmetic on "it" favors the pup, giving the second chain.

Code

The snippet below clusters mentions with a union-find pass over scored pairs, using only the standard library. Scores mirror the worked example.

1def cluster(mentions: list[str], pairs: dict[tuple[int, int], float], cutoff: float = 0.5) -> list[list[str]]:2    parent = list(range(len(mentions)))3    def find(x: int) -> int:4        while parent[x] != x:5            parent[x] = parent[parent[x]]6            x = parent[x]7        return x8    for (i, j), s in pairs.items():9        if s >= cutoff:10            parent[find(j)] = find(i)11    out: dict[int, list[str]] = {}12    for k, m in enumerate(mentions):13        out.setdefault(find(k), []).append(m)14    return sorted(out.values(), key=len, reverse=True)15ms = ["Mira", "pup", "She", "it"]16sc = {(0, 2): 0.74, (1, 3): 0.81, (0, 3): 0.12}17print(cluster(ms, sc))  # -> [['Mira', 'She'], ['pup', 'it']]

Watch Out For

The Quick Version

• Coreference resolution clusters mentions of one entity into a chain, so Mira, She and her resolve to a single person.
• The pipeline is mention detection, pairwise antecedent scoring with similarity plus distance and agreement features, then clustering.
• The score $s(i, j)$ blends span vectors, their product, and metadata; each mention links to its best antecedent or starts a new entity.
• In the worked example, She links to Mira with probability about 0.74 from a score of 1.64 against weaker rivals.
• Chains feed question answering, summarization and extraction; broken chains mean miscounted entities downstream.

What to Read Next

• Named Entity Recognition
• Relation Extraction
• Information Extraction