US 12,406,148 B2
Method, device, and system for tracking dialogue state
Jamin Shin, Seoul (KR); June Young Park, Yongin-si (KR); Han Gyeol Yu, Seoul (KR); and Hyeong Don Moon, Seoul (KR)
Assigned to RIIID INC., Seoul (KR)
Filed by RIIID INC., Seoul (KR)
Filed on Feb. 14, 2023, as Appl. No. 18/168,658.
Claims priority of application No. 10-2022-0018594 (KR), filed on Feb. 14, 2022.
Prior Publication US 2023/0259715 A1, Aug. 17, 2023
Int. Cl. G06F 40/35 (2020.01); G06F 16/34 (2019.01)
CPC G06F 40/35 (2020.01) [G06F 16/345 (2019.01)] 8 Claims
OG exemplary drawing
 
1. A method of tracking a dialogue state, which is performed by a device for tracking a dialogue state that acquires target dialogue data and generates a dialogue state template based on the target dialogue data, the method comprising:
acquiring a trained dialogue state tracking model;
acquiring the target dialogue data;
acquiring dialogue summary data from the target dialogue data using the dialogue state tracking model; and
generating the dialogue state template from the dialogue summary data,
wherein the dialogue state tracking model includes an input layer for receiving the target dialogue data, an output layer for outputting the dialogue summary data, and a hidden layer having a plurality of nodes connecting the input layer and the output layer, and is trained using a training set that includes dialogue data and a dialogue summary sentence generated from dialogue state data related to the dialogue data,
wherein the dialogue state tracking model is trained by generating the dialogue summary sentence based on the dialogue state data using a converter which is pre-trained language model (PLM), and the PLM is one of Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized BERT Approach (ROBERTa), Bayesian Additive Regression Trees (BART) and Sequence-to-Sequence (Seq2Seq),
wherein the dialogue state tracking model is configured to receive the dialogue data through the input layer and output a dialogue summary prediction value through the output layer during training, and is trained by updating a parameter of at least one node included in the dialogue state tracking model based on a similarity between the dialogue summary prediction value and the dialogue summary sentence, and
wherein the generating of the dialogue state template includes:
identifying a target sentence prefix included in the dialogue summary data,
determining a target domain related to the target summary data based on the identified target sentence prefix,
extracting at least one target sentence related to the target domain from among a plurality of sentences included in the dialogue summary data, and
generating the dialogue state template based on the extracted target sentence.