| 摘要 |
PROBLEM TO BE SOLVED: To detect a topic boundary at real time.SOLUTION: When an utterance A is inputted, topic boundary detection means 21 determines, with respect to a row Y of N pieces continuous utterances including the utterance A and utterances inputted before the input of the utterance A and a row X of continuous utterances of a same topic inputted immediately before the row Y of utterances, that a boundary between the row X of utterances and the row Y of utterance is a topic boundary when the dispersion of a sum aggregate U of an aggregate of concept vectors which is obtained by referring to a concept base 22 for each word in the X, and an aggregate of concept vectors which is obtained by referring to the concept base 22 for each word in the Y, is equal to or more than a threshold p, and when a value which is obtained by dividing a sum of variation in cluster which is calculated to the aggregate of the concept vectors for each word in the X and variation in cluster which is calculated to the aggregate of the concept vectors for each word in the Y, by variation in cluster which is calculated to the sum aggregate U is equal to or less than a threshold q. |
