Parameter learning apparatus, parameter learning method, and computer readable recording medium

A parameter learning method performed by a computer and comprising:

• extracting, from a set of entities that are each a string of words, a document in which the entities appear;
• extracting information regarding appearance positions of the entities in the document;
• extracting a pair of one entity that appears in a predetermined range in the document and a text representation including a string of words that is present in the predetermined range, and information indicating the appearance position of the one entity in the string of words included in the text representation, as a one-term document fact;
• obtaining, as a predicate fact, a plurality of entities in a predetermined range and a predicate relating to the entities, from a database in which entities and related predicates are registered in advance;
• outputting, for each of the plurality of entities that constitute the predicate fact, a one-term predicate that is defined from the predicate that constitutes the predicate fact and the one entity, as a one-term partial predicate fact, using the predicate fact;
• calculating a one-term partial predicate feature vector that represents a feature of the predicate in the one-term partial predicate fact, a one-term text representation feature vector that represents a feature of the text representation in the one-term document fact, and a one-term entity feature vector that represents a feature of one of the plurality of entities in the one-term partial predicate fact or the one entity in the one-term document fact, using a set parameter;
• calculating, as a first one-term score with respect to the one-term partial predicate fact, a score indicating a degree of establishment of the one-term partial predicate fact, using the one-term partial predicate feature vector and the one-term entity feature vector;
• calculating, as the first one-term score with respect to the one-term document fact, a score indicating a degree of establishment of the one-term document fact, using the one-term text representation feature vector and the one-term entity feature vector;
• calculating, as a second one-term score with respect to a combination of the one entity and a predicate that is not extracted as the one-term partial predicate fact, a score indicating a degree of establishment of the combination;
• calculating, as the second one-term score with respect to a combination of one entity and a text representation that is not extracted as the one-term document fact, a score indicating a degree of establishment of the combination;
• updating the parameter by using a gradient method, such that the first one-term score is higher than the second one-term score;
• calculating and outputting a score indicating a degree of establishment of the predicate fact using the first one-term score and the second one-term score as to the one-term partial predicate fact, using the updated parameter;
• and
• calculating and outputting a score indicating a degree of establishment of a combination of the plurality of entities and a predicate related to the entities that is not obtained as the predicate fact, using the updated parameter.

A non-transitory computer readable recording medium storing instructions that cause a computer to perform processing comprising:

• extracting, from a set of entities that are each a string of words, a document in which the entities appear;
• extracting information regarding appearance positions of the entities in the document;
• extracting a pair of one entity that appears in a predetermined range in the document and a text representation including a string of words that is present in the predetermined range, and information indicating the appearance position of the one entity in the string of words included in the text representation, as a one-term document fact;
• obtaining, as a predicate fact, a plurality of entities in a predetermined range and a predicate relating to the entities, from a database in which entities and related predicates are registered in advance;
• outputting, for each of the plurality of entities that constitute the predicate fact, a one-term predicate that is defined from the predicate that constitutes the predicate fact and the one entity, as a one-term partial predicate fact, using the predicate fact;
• calculating a one-term partial predicate feature vector that represents a feature of the predicate in the one-term partial predicate fact, a one-term text representation feature vector that represents a feature of the text representation in the one-term document fact, and a one-term entity feature vector that represents a feature of one of the plurality of entities in the one-term partial predicate fact or the one entity in the one-term document fact, using a set parameter;
• calculating, as a first one-term score with respect to the one-term partial predicate fact, a score indicating a degree of establishment of the one-term partial predicate fact, using the one-term partial predicate feature vector and the one-term entity feature vector;
• calculating, as the first one-term score with respect to the one-term document fact, a score indicating a degree of establishment of the one-term document fact, using the one-term text representation feature vector and the one-term entity feature vector;
• calculating, as a second one-term score with respect to a combination of the one entity and a predicate that is not extracted as the one-term partial predicate fact, a score indicating a degree of establishment of the combination;
• calculating, as the second one-term score with respect to a combination of one entity and a text representation that is not extracted as the one-term document fact, a score indicating a degree of establishment of the combination;
• updating the parameter by using a gradient method, such that the first one-term score is higher than the second one-term score;
• calculating and outputting a score indicating a degree of establishment of the predicate fact using the first one-term score and the second one-term score as to the one-term partial predicate fact, using the updated parameter;
• and
• calculating and outputting a score indicating a degree of establishment of a combination of the plurality of entities and a predicate related to the entities that is not obtained as the predicate fact, using the updated parameter.