US 11,941,365 B2
Response selecting apparatus, model learning apparatus, response selecting method, model learning method, and program
Ryuichiro Higashinaka, Tokyo (JP); Masahiro Mizukami, Tokyo (JP); and Junji Tomita, Tokyo (JP)
Assigned to NIPPON TELEGRAPH AND TELEPHONE CORPORATION, Tokyo (JP)
Appl. No. 17/050,347
Filed by NIPPON TELEGRAPH AND TELEPHONE CORPORATION, Tokyo (JP)
PCT Filed Apr. 9, 2019, PCT No. PCT/JP2019/015437
§ 371(c)(1), (2) Date Oct. 23, 2020,
PCT Pub. No. WO2019/208199, PCT Pub. Date Oct. 31, 2019.
Claims priority of application No. 2018-086370 (JP), filed on Apr. 27, 2018.
Prior Publication US 2021/0089723 A1, Mar. 25, 2021
Int. Cl. G06F 40/35 (2020.01); G06N 5/04 (2023.01); G06N 20/00 (2019.01)
CPC G06F 40/35 (2020.01) [G06N 5/04 (2013.01); G06N 20/00 (2019.01)] 15 Claims
OG exemplary drawing
 
1. A response selecting apparatus for outputting one or a plurality of responses to an input question, which is inputted a question, the response selecting apparatus comprising:
a recording medium recording question/answer pairs, which are pairs of a question and an answer, and a learned translation model for obtaining an answer from a question; and
processing circuitry configured to:
execute a document search processing which uses the input question as input, searches for a question/answer pair by a search method determined in advance from the question/answer pairs recorded in the recording medium, and outputs the searched question/answer pair as a search-result question/answer pair;
execute a score calculation processing which uses the input question and the answer of the search-result question/answer pair as input, obtains a translation likelihood, which is a numerical value based on a probability of obtaining the answer from the input question, by using the translation model, and calculates a score of each of the search-result question/answer pair with respect to the input question based on the translation likelihood; and
execute a ranking processing which selects a predetermined number of the search-result question/answer pair in an order of appropriateness indicated by the score and outputs the answer of the selected search-result question/answer pair as a response;
wherein
the translation model is learned in accordance with a degree indicated by an index by using a pair of a question and an answer associated with the index indicating the degree of how much the answer sounds like a character, and
the index indicating the degree of how much the answer sounds like the character includes: a first index evaluated by a person other than the character himself/herself, a second index indicating that the answer is responded by the character himself/herself, and a third index indicating that the character himself/herself certified that the respond sounds like the character.