MODEL LEARNING DEVICE, RANKING UNIT, METHOD AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To realize learning of a model for precisely ranking a speech string on the basis of cohesion of the speech string.SOLUTION: A pseudo negative example generation part 24, on the basis of a speech string composed of multiple continuous speeches that are input as a positive example, generates a speech string after the last speech included in the speech string is replaced with another speech, as a pseudo negative example. A feature quantity extraction part 28 extracts a feature quantity on the basis of a result of morphological analysis to the speech string input as the positive example, and extracts a feature quantity on the basis of a result of morphological analysis to the generated speech string. A model learning part 30, on the basis of the extracted feature quantities, learns a model for determining cohesion of the speech string so that the feature quantity of the speech string input as the positive example is ranked higher than the feature quantity of the generated speech string.