Method for constructing double-discriminator dialogue generation model based on generative adversarial network

Abstract

The invention discloses a method for constructing a double-discriminator dialogue generation model based on a generative adversarial network. The method comprises the following steps: firstly, processing corpora to obtain quaternary corpora with similar dialogue information; secondly, pre-training a rewriting model and a discrimination model, enabling the rewriting model to rewrite the matched similar reply so as to generate a reply more conforming to the current context, enabling the discrimination model to discriminate true and false statements, and discriminating whether the statements comefrom a corpus or the rewriting model; and finally, performing adversarial learning on the rewriting model and the discrimination model, and obtaining an optimal rewriting effect in the game process of the two models. According to the method, two discriminators are introduced to improve the generation model from multiple angles, and great progress is achieved in the aspects of syntax, context correlation and the like of generated sentences.

Description

technical field [0001] The invention relates to natural language processing, deep learning, and dialog systems, and relates to a method for constructing a dual-discriminator dialog generation model based on a generative confrontation network (GAN). Background technique [0002] With the development of smart phones and smart homes, the interaction between people and machines has become more frequent, and users have higher and higher requirements for the quality of dialogue with machines, hoping to get a smooth and diverse communication experience. It means that the template-based dialogue system commonly used in the industry is no longer able to meet the needs of users. At present, the methods commonly used in the industry to build dialogue systems are mostly based on templates, that is, manually sorting out and defining a large number of words to form templates, and inputting what users say into the pre-defined templates to get fixed replies. This method covers fewer topics,...

AI Technical Summary

Problems solved by technology

At present, the "self-study" of most generative dialogue systems is based on the seq2seq model, allowing the model to learn how to encode the input, and then decode it to get a reply, and improve the model by continuously narrowing the gap between the generated sentence and the real sentence. However, this It will cause the model to be "lazy" - it only learns simple generation, that is, it is more inclined to generate universal responses such as "I don't know", "OK", and "um".

At the same time, since the corpus used in these traditional methods is a binary corpus such as "above + reply", there is no intervention from the reply sentence during the generation process, so that the model does not know what kind of sentence is correct. As a result, some grammatical errors and semantically incomprehensible discourses are sometimes generated

Images