Method and device for tracking dialogue state in goal-oriented dialogue system

An apparatus for tracking a dialogue state, comprising:

• a processor;
• and memory storing instructions thereon, the instructions when executed by the processor cause the processor to:
• obtain target utterances between a dialogue system and a user;
• generate a first contextual semantic matrix by encoding respective words included in the target utterances and to calculate, by processing vectors constituting the first contextual semantic matrix based on a relevance of the vectors to a plurality of second contextual semantic vectors pre-stored and corresponding respectively to a plurality of domain-slot types, a context vector for each second contextual semantic vector;
• and
• input each context vector into a learning model that is pre-trained and to derive, from a plurality of third contextual semantic vectors pre-stored and corresponding respectively to a plurality of candidate slot values, a slot value corresponding to a third contextual semantic vector alone having a minimum distance from an output of the learning model for each context vector,
• wherein the context vector is calculated by calculating weights with a multi-head attention having parameters that are trained and by obtaining and using a scaled dot-product between each second contextual semantic vector and column vectors constituting the first contextual semantic matrix, and applying the weights to the column vectors constituting the first contextual semantic matrix to calculate a weighted sum of the column vectors.

A non-transitory computer-readable medium storing computer-executable instructions for causing, when executed by the computer, the computer to perform a method comprising:

• obtaining target utterances between a dialogue system and a user;
• generating a first contextual semantic matrix by encoding respective words included in the target utterances;
• calculating, by processing vectors constituting the first contextual semantic matrix based on a relevance of the vectors to a plurality of second contextual semantic vectors pre-stored and corresponding respectively to a plurality of domain-slot types, a context vector for each second contextual semantic vector;
• inputting each context vector into a learning model that is pre-trained;
• and
• deriving, from a plurality of third contextual semantic vectors pre-stored and corresponding respectively to a plurality of candidate slot values, a slot value corresponding to a third contextual semantic vector alone having a minimum distance from an output of the learning model for each context vector,
• wherein the calculating of the context vector comprises calculating weights with a multi-head attention having parameters that are trained and by obtaining and using a scaled dot-product between each second contextual semantic vector and column vectors constituting the first contextual semantic matrix, and applying the weights to the column vectors constituting the first contextual semantic matrix to calculate a weighted sum of the column vectors.