Research On Multi-Turn Dialogue Technology Of The Open Domain Based On Deep Learning

Dialogue system is an important and challenging task of artificial intelligence that investigates how to endow machines the ability to communicate with humans using natural language.In recent years,with the rapid growth of social media platforms such as Weibo and Twitter,researchers have gained easy access to a large amount of dialogue data,and building a data-driven open-domain dialogue system has become a hot research.Among them,the multi-turn dialogue system is in line with the nature of human dialogue and has a broad application prospect,so this paper mainly explores the open-domain multi-turn dialogue system.Although various models have been contributed by researchers to improve the performance of the dialogue system,there are still some problems to be solved in both the retrieval dialogue model and the generation dialogue model.In addition,it is worth exploring to combine retrieval and generation models to improve the quality of the response.Therefore,this paper conducts research on multi-turn dialogue technology from the following three aspects,and proposes specific solutions based on deep learning.(1)Research on the retrieval dialogue model.To solve the problem that the response retrieved by the retrieval dialogue model lacks context relevance,this paper proposes a context interaction deep matching network(CIMN)to match context and response.The model regards the dialogue context as consisting of multiple single-turn conversations,using the response of each single-turn conversation to measure the importance of each part in question,and after obtaining the weighted representation of all words and utterances in the context,then matching with the response.This is helpful to identify the semantic relationship between context and response,so as to select the context-sensitive response.(2)Research on generation dialogue model.To solve the problem that generation dialogue models tend to generate generic response,this paper proposes a hierarchical structured multi-head attention network(HMAN)to model the response generation process.The utterance encoder,context encoder and response decoder of this model are all based on the multi-head attention network,so that the encoded utterance representation and context representation can capture the complex dependencies within and between utterances,thus generating more diverse and informative responses.(3)Research on hybrid dialogue model combining retrieval and generation.To further improve the response quality,this paper proposes a hybrid dialogue model based on information fusion and reranking.The model inputs the retrieved candidate responses into the generation model to provide additional relevant supplementary information for the generation model;the generated response and the retrieved response are reranked to eliminate the retrieved irrelevant response and the generated meaningless response,so that the dialogue system can return the most appropriate response.