US 12,087,281 B2
Systems and methods for unsupervised structure extraction in task-oriented dialogues
Liang Qiu, San Jose, CA (US); Chien-Sheng Wu, Mountain View, CA (US); Wenhao Liu, Redwood City, CA (US); and Caiming Xiong, Menlo Park, CA (US)
Assigned to Salesforce, Inc., San Francisco, CA (US)
Filed by Salesforce, Inc., San Francisco, CA (US)
Filed on Jan. 31, 2022, as Appl. No. 17/589,693.
Claims priority of provisional application 63/256,190, filed on Oct. 15, 2021.
Prior Publication US 2023/0120940 A1, Apr. 20, 2023
Int. Cl. G10L 15/183 (2013.01); G10L 15/05 (2013.01); G10L 15/06 (2013.01); G06N 20/00 (2019.01)
CPC G10L 15/063 (2013.01) [G10L 15/05 (2013.01); G10L 15/183 (2013.01); G06N 20/00 (2019.01); G10L 2015/0631 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method of unsupervised dialogue structure extraction, the method comprising:
receiving a training corpus containing at least one dialogue which includes a plurality of conversational turns, at least one conversational turn including a system response and a user utterance, the user utterance including a plurality of tokens;
determining, based on a classification model, one or more contiguous spans over one or more subsets of the plurality of tokens;
generating token span encodings in a representation space based on the one or more contiguous spans and a context including the user utterance;
assigning the token span encodings into separate slot groups based on a clustering of the token span encodings in the representation space and a pre-defined number of slot groups;
generating a dialogue structure based on a sequence of states corresponding to predicted tags associated with the slot groups; and
incorporating the dialogue structure with the at least one dialogue as training data for an intelligent dialogue agent.