Response selecting apparatus, response selecting method, and response selecting program

A non-transitory computer-readable recording medium on which the response selecting program for causing a computer to operate as a response selecting apparatus is recorded, the response selecting apparatus comprising:

• a recording medium in which a plurality of question-answer pairs which are pairs of questions and answers are recorded;
• and
• processing circuitry configured to:
• execute a document searching processing which searches for a question-answer pair from the question-answer pairs recorded in the recording medium using the input question as input using a predetermined search scheme, and outputs the question-answer pair as a search-result-question-answer pair;
• execute an information acquiring processing which acquires information for quantifying appropriateness of the search-result-question-answer pair with respect to the input question using the input question and the search-result-question-answer pair as input, and outputs the information as quantification information;
• execute a score calculating processing which calculates a score with respect to the input question for each of the search-result-question-answer pairs from a numerical value indicating appropriateness based on the search scheme and a numerical value based on the quantification information;
• and
• execute a ranking processing which selects the search-result-question-answer pairs of a predetermined number in descending order of appropriateness indicated by the scores, and outputs answers of the selected search-result-question-answer pairs as responses,
• wherein the information acquiring processing, as the quantification information,
• estimates information for specifying a question type of the input question and a question type of the search-result-question-answer pair,
• extracts word strings which become topics as focus words for the input question and the question of the search-result-question-answer pair,
• outputs a numerical value based on a probability of the answer being obtained from the input question as a translation likelihood and a numerical value based on a probability of the input question being obtained from the answer as a reverse translation likelihood using the input question and the answer of the search-result-question-answer pair as input, and
• obtains utterance vectors indicating a semantic content for the input question and the question of the search-result-question-answer pair;
• and
• the score calculating processing calculates a score(Q, (Q′, A′)) as
• score(Q,(Q′,A′))=w1·search_score+w2·qtypes_match_score+w3·center-word_score+w4·translation_score+w5·rev_translation_score+w6·semantic_similarity_score
• where Q is the input question, Q′ is the question of the search-result-question-answer pair, A′ is the answers,
• search_score is a numerical value indicating appropriateness based on the search scheme,
• qtypes_match_score is a numerical value indicating higher appropriateness as the question type of the input question better matches the question type of the search-result-question-answer pair based on the information for specifying the question type of the input question and the question type of the search-result-question-answer pair,
• center-word_score is a numerical value indicating higher appropriateness as the focus word of the input question better matches the focus word of the question of the search-result-question-answer pair,
• translation_score is a numerical value indicating higher appropriateness as the translation likelihood is higher,
• rev_translation_score is a numerical value indicating higher appropriateness as the reverse translation likelihood is higher,
• semantic_similarity_score is a numerical value indicating higher appropriateness as the utterance vector of the input question is more similar to the utterance vector of the question of the search-result-question-answer pair, and
• w1, . . . , w6 are weight for determining a numerical value on which emphasis is to be placed.