Multi-label entities: use max similarity across all names during propagation

[monneyboi]

Problem

After a merge, a merged entity has multiple names in its occurrence pool (e.g. "Meridian Technologies" and "Meridian Tech"), but only the most-frequent canonical name is stored on Entity.name and embedded. The other names are silently dropped.

In prepare_embeddings, only the canonical name is embedded:

all_names = sorted({e.name for g in graphs for e in g.entities.values()})

In propagate, similarity is initialised from that single name:

name_a = graph_a.entities[eid_a].name
emb_a = name_embeddings.get(name_a)

This means if an already-merged entity's canonical name is not the closest match to a name in another graph, similarity is under-estimated.

Correct behaviour

For entities with multiple known names, the initial similarity between two entities should be:

max(cosine_sim(emb_a_i, emb_b_j) for emb_a_i in all_embs_a for emb_b_j in all_embs_b)

Fix

1. Add a names: set[str] field to Entity (alongside the canonical name).
2. Populate it from all occurrence names at load/merge time.
3. In prepare_embeddings, embed all names (not just canonical ones).
4. In propagate, compute initial sigma as max pairwise similarity across all name embeddings for each entity pair.

When this matters

The current pipeline runs one pass of matching on the original per-article graphs (which are always single-source, so single-name). The bug only bites if merged graphs are fed back into a second matching pass (iterative refinement). It's latent today but will silently degrade quality if iterative matching is added.

[monneyboi]

@claude this is a old issue. Is this still a thing?

[claude]

Claude finished @monneyboi's task in 40s —— View job

---

Analysis

Yes, this issue is still valid — the underlying bug is still present, though the code has been refactored since the issue was written (the Entity class was renamed to Node, entities dict to nodes, etc.).

Current state:

graph.py — Node has only a single name field, no names: set[str]:

@dataclass
class Node:
    id: str
    graph_id: str
    name: str   # ← only one name, no multi-name support

match.py — name similarity in propagate_similarity uses only the single canonical name:

name_sim = max(
    0.0,
    soft_tfidf(graph.nodes[id_a].name, graph.nodes[id_b].name, idf),
)

match_graphs also only collects canonical names for IDF:

all_names = [node.name for graph in graphs for node in graph.nodes.values()]

Verdict: The issue description's diagnosis is still accurate, just with updated class/field names. The fix described in the issue (names: set[str] on Node, embed all names, take max pairwise similarity) is still the right approach.

As noted in the issue, the bug is latent — it won't affect the current single-pass pipeline (which only ever sees single-source graphs), but will silently degrade quality if iterative re-matching of merged graphs is added.