Training method and device of dialogue response generation model and dialogue response generation method

Abstract

The invention discloses a dialogue response generation model training method and device and a dialogue response generation method.The training method comprises the steps that preset sample data and a document library are obtained, and the sample data comprise dialogue process data and a correct document label and a topic keeping label of each round of dialogue; and traversing each round of dialogue corresponding to the dialogue process data by utilizing a dialogue response generation model, generating response information of the round of dialogue based on dialogue historical data generated before the response information of the round of dialogue and the document library, calculating a loss function value based on the corresponding label in the sample data, and calculating the corresponding label in the sample data by utilizing the loss function value. Parameters of the dialogue response generation model are optimized and adjusted; wherein during generation, the most relevant document used for generating the response information is selected from the document library on the basis of the current topic keeping condition in a secondary correlation mode. By adopting the method, the dialogue reply quality can be improved.

Description

technical field [0001] The invention relates to man-machine dialogue technology, in particular to a training method and device for a dialogue response generation model and a dialogue response generation method. Background technique [0002] Human-computer dialogue system is an important research direction in the field of natural language processing. With the development of deep neural network technology, systems such as Alexa and XiaoIce can have smooth dialogues with people in some fields. In order to make the dialogue system adapt to more fields, inspired by the dialogue of everyone, researchers began to explore adding external knowledge to the dialogue system. Among these external knowledge, documents are readily available and informative, and document-driven dialogue tasks are proposed. In order to make the dialogue more similar to the real dialogue between people, high-quality responses need to meet the requirements of language fluency, contextual relevance, and the ab...

AI Technical Summary

Problems solved by technology

" such generic responses that neither correctly reflect document knowledge nor satisfy context-sensitive requirements

[0008] In addition, the existing method of directly generating a reply encodes and integrates the dialogue context information and the document selected in the first stage, and then directly transfers the integration result to the decoder to generate a reply. In this way, the generated reply often lacks logic sex, so it is easy to have the problem of unsound sentences

Images