We achieve 94.88% accuracy on CoNLL (YAGO)⁠(opens in a new window) (previous state of the arts: 91.50⁠(opens in a new window)%, 91.70⁠(opens in a new window)%) and 90.85% on TAC KBP 2010 challenge⁠(opens in a new window) (previous state of the arts: 87.20⁠(opens in a new window)%, and 87.70⁠(opens in a new window)%). Previous methods used distributed representations⁠(opens in a new window). Types can go almost all the way on these tasks, as perfect type prediction would give accuracies of 98.6-99%.

Wikidata’s knowledge graph can be turned into a source of training data for mapping fine-grained entities to types. We apply its instance of⁠(opens in a new window) relation recursively to determine the set of types for any given entity — for example, any descendent node of the human⁠(opens in a new window) node has type human. Wikipedia can also provide entity-to-type mapping through its category link⁠(opens in a new window).

Wikipedia-internal link statistics provide a good estimate of the chance a particular phrase refers to some entity. However, this is noisy since Wikipedia will often link to specific instance of a type rather than the type itself (anaphora⁠(opens in a new window) — e.g. king → Charles I of England) or link from a nickname (metonymy⁠(opens in a new window)). This results in an explosion of associated entities (e.g. king has 974 associated entities) and distorted link frequencies (e.g. queen links to the band Queen⁠(opens in a new window) 4920 times, Elizabeth II⁠(opens in a new window) 1430 times, and monarch⁠(opens in a new window) only 32 times).

The easiest approach is to prune⁠(opens in a new window) rare⁠(opens in a new window) links⁠(opens in a new window), but this loses information. We instead use the Wikidata property graph to heuristically turn links into their “generic” meaning, as illustrated below.