Entity Linking and Knowledge Graph Construction › Summary

Entity linking grounds mentions to identifiers in a chosen KB snapshot. Separate mention detection, candidate retrieval, ranking, and handling of uncertainty or KB absence. For fixed gold in-KB mentions, candidate-constrained ranking cannot recover an entity excluded by retrieval. A nonentity span, an uncertain link, and an out-of-KB referent need different treatment.

The example combines an alias prior $P(e \mid m)$ with context-description similarity. Each signal can mislead. For an evaluated system, select the weights on development data and check the feature scales; the toy weights are illustrative. Their weighted score is not a calibrated probability.

Pairwise coherence can couple mention assignments when its weighted rewards cannot be reduced to separate per-mention scores. The toy uses a nonnegative reward derived from normalized inlink overlap. Its exhaustive search finds the highest-scoring assignment for each tiny document and changes the d8 prediction from the local winner to the assigned planet label. Strong coherence can also favor the wrong topic, so this constructed improvement does not establish a general accuracy gain.

Neural approaches include joint span-and-entity models, bi-encoder retrieval with cross-encoder reranking, and autoregressive entity-name retrieval. Description-based models can represent entities absent from supervised linker training. Whether an index update also needs retraining depends on the architecture. A generative output restricted to valid names can still identify the wrong referent.

Graph construction combines linked arguments with extracted candidate relations. Consistent identifiers support joins; provenance, temporal scope, and merge history support inspection and repair. RDF and property graphs offer different representations. The executable statement store illustrates assertion identity and provenance using Python collections; it does not implement RDF or SPARQL semantics.

Evaluation needs an explicit span-matching rule, NIL and abstention conventions, averaging scheme, KB version, and split policy. Measure the stages separately and include representative target-domain data. The worked example's seven, eleven, and twelve correct links explain the mechanics of three scores; they do not replace that evaluation.