Discovering types for entity disambiguation

Researchers introduced a type-based entity disambiguation approach that asks category questions instead of resolving ambiguity directly.

They trained a bidirectional LSTM on Wikipedia-linked data to predict type membership for mentions and candidate entities.

Candidate entities are then ranked using type predictions plus Wikipedia link frequency, improving speed and accuracy.

The method avoids expensive document-level coherence inference, making lookup closer to linear time and easier to extend to new entities.