US 12,277,121 B2
Data augmentation system, data augmentation method, and information storage medium
Chen Zhao, Tokyo (JP); Yuki Nakayama, Tokyo (JP); and Koji Murakami, Eastchester, NY (US)
Assigned to RAKUTEN GROUP, INC., Tokyo (JP)
Filed by RAKUTEN GROUP, INC., Tokyo (JP)
Filed on Jun. 19, 2023, as Appl. No. 18/337,061.
Claims priority of provisional application 63/367,323, filed on Jun. 29, 2022.
Prior Publication US 2024/0004880 A1, Jan. 4, 2024
Int. Cl. G06F 16/2455 (2019.01); G06F 16/2453 (2019.01); G06N 3/0442 (2023.01); G06N 3/0455 (2023.01)
CPC G06F 16/2455 (2019.01) [G06F 16/24542 (2019.01); G06N 3/0442 (2023.01); G06N 3/0455 (2023.01)] 19 Claims
OG exemplary drawing
 
1. A data augmentation system, comprising at least one processor configured to:
acquire a first search query including a first named entity, which was actually input in a search executed in a past;
execute training of a first model that outputs a second search query including a virtual second named entity, which is different from the first named entity, based on the first search query;
execute data augmentation based on the second search query output by the trained first model;
wherein the first model comprises a model having output determined in one of a stochastic manner or a random manner,
wherein the at least one processor is configured to execute the data augmentation based on the second search query output in one of the stochastic manner or the random manner by the trained first model;
wherein the first model comprises a variational auto-encoder (VAE) model having output determined in one of the stochastic manner or the random manner with a standard normal distribution followed by a probability density function to which a latent variable conforms,
wherein the at least one processor is configured to execute the data augmentation based on the second search query output in one of the stochastic manner or the random manner by the trained VAE model;
wherein the VAE model includes:
an encoder configured to learn mapping from the first search query input during the training to an array of a cell state of a long short-term memory (LSTM) assumed to be a normal distribution;
a decoder configured to learn mapping from an average value of the cell state in the array to a third search query output during the training,
wherein the first model is configured to cause the decoder to generate the second search query by setting the average value in the array sampled from the normal distribution as an initial cell state; and
wherein the second search query includes a fictitious word or phrase which does not exist in reality.