Emotion reply automatic generation method for open domain dialogue system

Abstract

The invention discloses an emotion reply automatic generation method for an open domain dialogue system, and aims to enable the dialogue system to automatically generate replies rich in content and appropriate in emotion according to input statements of a user. The method comprises the following steps: firstly, preprocessing input and reply statements of each sample in a corpus based on a word embedding table and a VAD sentiment dictionary to obtain an input and reply word embedding sequence and a sentiment embedding sequence; secondly, splicing word embedding and emotion embedding to expand word embedding, introducing replied emotion distribution information into the model, and encoding a spliced sequence in an encoder to obtain input and replied semantic representation vectors containing emotion information; and finally, predicting approximation of conditional prior distribution and posteriori distribution by using semantic representation vectors, sampling latent variables, and reconstructing replies and emotion distribution of the replies through the latent variables. When reply statements are generated, multiple replies are generated through multiple times of sampling, then the sequences are combined with the sequence model and the VAD dictionary, automatic scoring is conducted from the perspective of grammar, semantics and emotion of the replies, and the optimal reply isselected.

Description

technical field [0001] The invention belongs to the field of human-computer dialogue of natural language processing, and in particular relates to an automatic generation method of an emotional reply of an open domain dialogue system based on a deep neural network. Background technique [0002] In recent years, conversational robots have gradually integrated into people's daily life. For example, voice assistants such as Siri, Xiaodu, and Tmall Genie can help people complete some simple tasks through dialogue, such as setting alarm clocks, playing music, and chatting. At present, the existing dialogue systems can be divided into domain-specific dialogue systems and open-domain dialogue systems according to domain knowledge. Domain-specific dialogue systems are limited to background knowledge in a specific domain to conduct dialogues, and related to them are task-oriented dialogue systems, such as movie ticket purchase assistant robots. The open-domain dialogue system has no...

AI Technical Summary

Problems solved by technology

However, one problem with increasing diversity is that contextual relevance decreases

In addition, traditional pre-trained word vectors are usually trained by language models and do not contain any emotional information, which may easily cause the content generated by the model to be inappropriate in terms of emotional expression.

Images