Relation extraction using QandA
Embodiments of the present invention disclose a method, a computer program product, and a computer system for extracting natural language relations between entities. A computer receives a configuration for associating one or more natural language questions with one or more entities and identifies the one or more entities annotated within a document. The computer answers the natural language questions associated with the identified one or more entities based on context surrounding the identified one or more entities. The computer may further transmit the natural language questions associated with the identified one or more entities and the surrounding context to a question and answer service, then receive answers to the natural language questions from the question and answer service. The computer may further determine whether the received answers correctly describe the relation between the identified one or more entities and other entities within the extracted surrounding context.
1. A computer-implemented method of extracting entity relations, the method comprising:
associating, by a computer, one or more preprogrammed questions with one or more first entity types;
associating, by the computer, one or more second entity types with one or more answers to the one or more preprogrammed questions;
identifying, by the computer, an entity annotated within a document;
extracting, by the computer, a portion of content in a proximity to the entity;
determining, by the computer, whether the entity corresponds to at least one of the one or more first entity types;
based on determining that the entity corresponds to the at least one of the one or more first entity types, determining, by the computer, the one or more answers to the one or more questions based on the extracted portion of content, wherein the determined one or more answers describe a relation between the identified entity and one or more other entities included within the portion of content;
weighting, by the computer, the determined one or more answers;
ranking, by the computer, the determined one or more answers based on the weighting;
determining, by the computer, whether a first ranked answer of the determined one or more answers is correct by comparing an entity type corresponding to the first ranked answer to the one or more second entity types associated with the determined one or more answers to the one or more questions;
based on determining that the first ranked answer is incorrect, rewording, by the computer, the one or more questions;
determining, by the computer, one or more second answers to the one or more reworded questions based on the extracted portion of content; and
associating, by the computer, the one or more second answers to the one or more reworded questions with the entity.