US 12,292,904 B2
Text generation apparatus, text generation method, text generation learning apparatus, text generation learning method and program
Itsumi Saito, Tokyo (JP); Kyosuke Nishida, Tokyo (JP); Atsushi Otsuka, Tokyo (JP); Kosuke Nishida, Tokyo (JP); Hisako Asano, Tokyo (JP); and Junji Tomita, Tokyo (JP)
Assigned to NIPPON TELEGRAPH AND TELEPHONE CORPORATION, Tokyo (JP)
Appl. No. 17/435,022
Filed by NIPPON TELEGRAPH AND TELEPHONE CORPORATION, Tokyo (JP)
PCT Filed Feb. 25, 2020, PCT No. PCT/JP2020/007343
§ 371(c)(1), (2) Date Aug. 30, 2021,
PCT Pub. No. WO2020/179530, PCT Pub. Date Sep. 10, 2020.
Claims priority of application No. 2019-037616 (JP), filed on Mar. 1, 2019.
Prior Publication US 2022/0138239 A1, May 5, 2022
Int. Cl. G06F 16/332 (2019.01); G06F 40/205 (2020.01); G06F 40/30 (2020.01); G06F 40/56 (2020.01)
CPC G06F 16/3322 (2019.01) [G06F 40/205 (2020.01); G06F 40/30 (2020.01); G06F 40/56 (2020.01)] 12 Claims
OG exemplary drawing
 
1. A sentence generation device comprising a processor configured to execute operations comprising:
receiving input of a first sentence and an output length of a word string, wherein the output length of a word string indicates a target number of words in a second sentence, and the output length of the word string represents a condition of generating the second sentence from the first sentence;
estimating an importance score of each word in the first sentence using a pre-trained content selection model, wherein the importance score indicates a level of importance for determining whether to select said each word for generating the second sentence having the output length of the word string, the pre-trained content selection model determines the importance score based on a combination of attention match between the first sentence and the second sentence and semantic fusion between the first sentence and the second sentence, and the pre-trained content selection model comprises the output length of the word string in embedding form; and
generating the second sentence having the output length of the word string based on the importance score according to content selection by a combination of the pre-trained content selection model and sentence generation by a generative model.