Training method and device as well as computer equipment and readable storage medium

Abstract

The invention provides a training method and device as well as computer equipment and a readable storage medium. The training method is applied to a dialogue generating network, the dialogue generating network comprises a first generator and a second generator which have duality relation and also comprises a first discriminator and a second discriminator; and the training method comprises the following steps: performing quality discrimination on a first text, output by the first generator, by adopting the first discriminator, and obtaining first discrimination information, wherein the first text is output by the first generator according to a preceding part of a dialogue and used for replying the preceding part of the dialogue; performing the quality discrimination on a second text, outputby the second generator, by adopting the second discriminator, and obtaining second discrimination information, wherein the second text is output by the second generator according to a following partof the dialogue and used for being continued as a preceding part in front of the following part of the dialogue; and under constraint of the duality relation, training the first generator according to the first discrimination information, and training the second generator according to the second discrimination information. The training method provided by the invention can reduce training difficulty and improve training efficiency.

Description

Technical field [0001] This application relates to the field of artificial intelligence technology, in particular to a training method, device, computer equipment and readable storage medium. Background technique [0002] The general dialogue system can be applied to chat robots, mobile phone assistants, etc. As an important scene of artificial intelligence, it has received more and more attention. At present, the dialogue generation network in the general dialogue system can be trained based on the retrieval formula and the generative formula. Among them, the retrieval style requires a lot of manual assistance, the efficiency of training is low, and the retrieval style is limited by the corpus quality, resulting in low accuracy of the output of the general dialogue system. Therefore, based on the generative formula, more and more attention is paid to the training of the dialogue generation network in the general dialogue system. [0003] However, based on generative training, th...

AI Technical Summary

Problems solved by technology

[0005] This application proposes a training method, device, computer equipment, and readable storage medium, which are used to solve the problem of introducing the extracted topics or keywords into the dialog context generation process in the prior art, which is difficult and the maximum The calculation process of mutual information is extremely time-consuming; through the method of confrontational learning, the dialogue generation network is trained, the stability is low, and the diversity of the output of the general dialogue system has not been significantly improved.

Images