Dialogue generation method and device applied to multi-round dialogue system, equipment and medium

Abstract

The invention discloses a dialogue generation method and device applied to a multi-round dialogue system, equipment and a medium, and the method comprises the steps: obtaining a first round of dialogue of a user, carrying out the feature extraction of the first round of dialogue through a semantic understanding model, and obtaining the semantic feature representation of the content of the first round of dialogue; deep semantic information is extracted from the semantic feature representation by using an encoder, and a first round of intermediate variable is generated; extracting the first round of intermediate variables by using an intention feature calculation network to obtain intention feature representation; respectively decoding the intention feature representation and the first round of intermediate variable, inputting a decoding result into the generative network, completing response by the generative network, generating a response text of the first round of dialogue content, and returning the response text to complete the generation of the first round of dialogue; and waiting and obtaining the next round of conversation, and repeating the steps until the user triggers the ending condition, so that the multi-round conversation is ended. The method is suitable for a dialogue scene of multi-round interaction, response generation is achieved based on the current dialogue content and the historical dialogue records, and multi-round interaction is completed.

Description

technical field [0001] The invention relates to the field of dialogue systems, belongs to the natural language processing technology, and in particular relates to a dialogue generation method, device, equipment and medium applied to a multi-round dialogue system. Background technique [0002] Dialogue-based systems in the prior art mainly include the following two categories: [0003] The first type is a dialogue system based on matching or retrieval, patent application numbers CN113806509A, CN113901188A, methods based on matching or retrieval: the prior art relies on an almost completely closed domain knowledge base. The mechanism matches similar candidate sentences in the knowledge base, and finally outputs the reply sentence corresponding to the candidate sentence with the highest similarity through similarity calculation. Although this type of method has low implementation cost and strong operability, it is easily limited by the knowledge base in practical applications....

AI Technical Summary

Problems solved by technology

But on the other hand, limited by the size of the model, the existing related technologies are only applied to a single round of dialogue, and the multi-round dialogue with greater actual demand has less practical application.

[0005] Therefore, the existing methods based on matching or retrieval have a strong dependence on the domain knowledge base, which is usually a closed domain knowledge base. The limitation is that when encountering knowledge outside the domain, the quality of the response text given by the system is greatly reduced. , and its response library is relatively fixed, lacks flexibility, cannot play a guiding role in the dialogue, and is difficult to complete task-oriented dialogues, so most of them are only used in single-round dialogue scenarios

Although the generation-based method can avoid the strong dependence on the knowledge base, the training of the model in the open knowledge field often requires extremely large data sets, and the implementation cost and speed are much higher than those of the retrieval system. On the other hand, the generative system has a larger task space and lacks control over the overall intention and historical information during the dialogue interaction process, which leads to logic confusion and mismatched dialogue content in the interaction.

Images