Open domain dialogue generation method based on multi-granularity feature decoupling

Abstract

The invention discloses an open domain dialogue generation method based on multi-granularity feature decoupling. The open domain dialogue generation method comprises the steps of obtaining a data set and a dialogue category; randomly selecting a plurality of first training pairs in the data set to input a conditional variation auto-encoder CVAE model, and obtaining prior Gaussian distribution corresponding to the dialogue category; inputting a second training pair in the data set into an advanced conditional variation auto-encoder A-CVAE model to obtain reconstruction expectation loss and KL divergence; according to the reconstruction expectation loss, the KL divergence and the prior Gaussian distribution corresponding to the first training pair, obtaining total loss; performing reverse gradient optimization on the A-CVAE model according to the total loss to obtain a trained A-CVAE model; obtaining a to-be-replied dialogue question; and inputting the to-be-replied dialogue question into the training A-CVAE model to generate a dialogue reply. The technical problems that an existing model cannot well capture category semantics of dialogues and the generated dialogues are not high in interpretability can be solved.

Description

technical field [0001] The invention relates to the technical field of man-machine dialogue, in particular to an open domain dialogue generation method based on multi-granularity feature decoupling. Background technique [0002] Human-computer dialogue is mainly divided into task-oriented and non-task-oriented dialogue application systems. Although these two deep learning black-box models of human-computer dialogue have made continuous progress in dialogue quality, the end-to-end model still lacks interpretability, suggesting that the model is programmatically unverifiable and incomprehensible. In response to the above problems, the prior art proposes to use a deep hidden variable model to improve the interpretability of dialogue generation by decoupling hidden variables into different feature representations. However, because the above-mentioned dialogue model is unsupervised training, it cannot effectively capture the specific semantic information of the dialogue. Therefo...

AI Technical Summary

Problems solved by technology

[0003] The main purpose of the present invention is to propose an open domain dialogue generation method based on multi-granularity feature decoupling, aiming to solve the technical problem that the existing models cannot capture the category semantics of the dialogue well, and the interpretability of the generated dialogue is not high

Images