CanonWeave: Preventing Commitment Interference in Partial-Ontology Schema Resolution
Abstract
city and product_name can both resemble generic string fields to a local scorer, yet linking them forces incompatible ontology types and changes the feasible matches and blocking keys of their component. CanonWeave treats this commitment interference as a state-conditioned scheduling problem: the early and late policies jointly change prompt type context, queried edges, bundle volume, and solver initialization, while the method itself coordinates types and links before impact-aware LLM arbitration. Across six public data lakes, 27.7% of pre-resolution commitments enter a constraint conflict, and conflict/propagation increase from 15.2% / 1.4 in sparse components to 45.7% / 4.2 in dense ones, defining a verifier-placement phase map over coupling strength. The complete same-graph policy improves Match F1 by 3.5 points and ER F1 by 2.6 points while requesting 211 fewer bundles per lake; against prompt-first LLM-Chain, CanonWeave gains 5.1–9.7 Match-F1 and 3.2–5.2 ER-F1 points as mean FD/unit/temporal violations fall from 5.1/2.4/3.5% to 1.7/1.0/1.5%. Fixing queried edges and verifier outputs retains a 2.0-point gain from the coordinated initializer, isolating path dependence under identical evidence; a second resolver and Qwen2.5-72B preserve the state-conditioned advantage. The resulting design rule for partial-ontology lakes is direct: coordinate the canonical state before expensive local evidence acquires authority.
est. 32% chance this paper gets accepted at ICLR 2027.
What do you think this paper will get?
All positions stay anonymous.