📄 Relational Inference for Wikification

Summary

Wikification (D2W) links mentions in a document to the most specific Wikipedia page, or to NIL when no page fits. Systems that rely on statistics such as Pr(title|surface) and context coherence still make errors that any human reader would avoid. For example, they collapse "Mubarak" (the wife) into Hosni Mubarak, or they never consider Socialist Party of Serbia as a candidate for "Socialist Party". Cheng and Roth add relational analysis on top of an existing Wikifier (GLOW). They extract high-precision textual relations, namely coreference and local syntactico-semantic relations such as apposition and possessives. They then check candidate title pairs against relation triples in an external knowledge base. The resulting evidence enters a Constrained Conditional Model solved as a sparse ILP. Relations are used both to add new candidates and to rerank them globally. The authors report clear gains over top Wikification systems on standard datasets. Without any training on TAC data, the system also matches the top tier of the TAC 2011 Entity Linking task. The work shows that light text understanding combined with background knowledge fixes errors that statistics alone cannot.

Key points

• The pipeline has five stages:
  • generate mentions;
  • get candidate priors from an existing Wikifier;
  • add non-coreference relational constraints and the candidates they suggest;
  • add coreference constraints and candidates;
  • solve the ILP for the best mention-to-title mapping.
• The ILP objective adds up two kinds of score:
  • a prior score for each mention-title choice;
  • pairwise relation weights for title pairs selected together.
• Each relation weight combines the relation's type and confidence from the text with the confidence of a matching triple in the KB.
• Constraints force exactly one title per mention. Because most pairwise weights are zero, the program stays small and Gurobi solves it quickly.
• On TAC 2011 Entity Linking (official modified B³ and Micro-Average):
  • RI reaches 86.1 MA and 83.7 B³ F1.
  • The best TAC systems are LCC (84.6 F1) and MS-MLI (86.8 MA).
  • The original GLOW-based Cogcomp entry scored 76.1 F1.
  • A baseline with no inference (RI-0), using an updated Wikipedia index and fuzzy lexical matching, gains 2.7 F1 over Cogcomp.
  • Full relational inference adds a further 4.9 F1.
• Unlike LCC, MS-MLI and NUSchime, RI was not trained on TAC data and used no query expansion. NIL clustering was trivial: entities were grouped by a shared out-of-KB URL or an identical surface form.

Interesting ideas

• I1. Relations as candidate generators, not only rerankers. A relation found in the text, such as a possessive link to a known person, can be used to query the KB. This can bring in a title that frequency-based lookup would bury or miss. The idea helps any linking task with a long tail of rare or never-linked names.
• I2. Inference-time modularity. The relational layer sits on top of any Wikifier that outputs a distribution over candidates per mention. The base system can be swapped without retraining, which is convenient when adapting to a new language or a different KB.
• I3. Precision over recall for constraints. The authors found that treating every context-window co-occurrence as a constraint created many false positives and hurt results. Restricting constraints to high-precision relations is a good general rule for adding soft structure to joint models.
• I4. Sparse pairwise ILP. Pair variables are created only where a relation weight is nonzero. This keeps global document-level inference tractable with off-the-shelf solvers, and the same trick transfers to other joint assignment problems such as coreference or relation extraction.
• I5. Cross-task evaluation without retraining. Running the system unchanged on TAC Entity Linking, with a separate ablation run (RI-0), separates robustness from tuning on development data. This is a useful way to make claims about generalization credible.