INFERENCE DEVICE, INFERENCE METHOD, AND INFERENCE PROGRAM

WIPO Patent Application WO/2021/010203

A1

The present invention improves the performance of a learned neural network that uses positional information indicating in which position in an input sequence each token included in the input sequence exists. The inference device includes a first generation unit that generates an intermediate sentence expression on the basis of a first sentence expression having information indicating the value of each token included in an input sequence and first positional information indicating in which position in the input sequence each token exists, a second generation unit that generates second positional information by modifying the first positional information on the basis of the first sentence expression and the intermediate sentence expression, and generates a hidden state expression on the basis of the second positional information and the intermediate sentence expression, and a third generation unit that generates a second sentence expression on the basis of the intermediate sentence expression and the hidden state expression.

CHEN KEHAI (JP)
WANG RUI (JP)
UCHIYAMA MASAO (JP)
SUMITA EIICHIRO (JP)

PCT/JP2020/026397

January 21, 2021

July 06, 2020

Click for automatic bibliography generation  

NAT INST INF & COMM TECH (JP)

G06N3/02; G06F40/44

NING DAI ET AL.: "Style Transformer: Unpaired Text Style Transfer witheout Disentangled Latent Representation", 21 June 2019 (2019-06-21), XP081371016, Retrieved from the Internet
KAWARA, YUKI ET AL.: "Analysis of the effects of presorting in neural machine translation", PROCEEDINGS OF THE 25TH ANNUAL MEETING OF THE ASSOCIATION FOR NATURAL LANGUAGE PROCESSING, 4 March 2019 (2019-03-04), pages 1455 - 1458

FUKAMI PATENT OFFICE, P.C. (JP)

View/Download PDF      PDF Help

Previous Patent: POSITION DETECTION SYSTEM

Next Patent: RESIN SHEET FOR SEALING