Posterior probability pursuit for entity disambiguation

A system that disambiguates a mention of an ambiguous entity, comprising:

• at least one processor;
• and
• memory that comprises computer-executable instructions that, when executed by the at least one processor, cause the at least one processor to perform acts including:
• receiving a document that comprises the mention of the ambiguous entity;
• identifying a set of candidate entities from an entity knowledge base from which to disambiguate the mention of the ambiguous entity;
• retrieving the set of the candidate entities from the entity knowledge base, wherein each of the candidate entities has a respective entity feature representation;
• forming a document feature representation based upon features of the document and the respective entity feature representations of the candidate entities, the document feature representation being formed to comprise features that are part of both:
• the features of the document;
• and
• a union set of features from the entity feature representations of the candidate entities;
• selecting a subset of features from the document feature representation by employing a posterior probability pursuit algorithm;
• and
• determining a disambiguated result for the mention of the ambiguous entity based upon the subset of the features by employing the posterior probability pursuit algorithm, wherein the disambiguated result is one of:
• an unknown entity;
• or
• one of the candidate entities.

A method of disambiguating a mention of an ambiguous entity, comprising:

• receiving a document that comprises the mention of the ambiguous entity;
• retrieving a set of candidate entities from an entity knowledge base based upon the mention of the ambiguous entity, wherein each of the candidate entities has a respective entity feature representation;
• generating a document feature representation based upon features of the document and the respective entity feature representations of the candidate entities, wherein features from the document feature representation comprise words and phrases;
• while conditions remain unsatisfied, iteratively:
• identifying a feature from the document feature representation that maximizes a posterior probability, wherein the feature is identified from remaining features of the document feature representation;
• and
• adding the feature to a subset of the features from the document feature representation;
• responsive to at least one of the conditions being satisfied:
• computing, based upon the subset of the features, respective posterior probabilities for the candidate entities and an unknown entity;
• and
• identifying a disambiguated result for the mention of the ambiguous entity, wherein the disambiguated result corresponds to a maximum posterior probability from the respective posterior probabilities, wherein the disambiguated result is one of:
• an unknown entity;
• or
• one of the candidate entities.