| In recent years,the massive available session data and deep learning methods have flourished,and the open domain multi-round session system has become a major trend and field.Building a natural and meaningful intelligent conversation system is a long-term goal in the field of artificial intelligence.Due to the existence of large-scale corpus,researchers have tried two routes.One is to search in the corpus and use the obtained content as a reply.However,the existing corpus is too large to carry out efficient and accurate retrieval;The second is to use the technology of text generation to generate new conversation content as replies,but the traditional generative conversation lacks common sense and tends to generate replies with grammar problems.Therefore,in order to solve and improve the above problems,this paper mainly studies the conversation content generation model using the hybrid method to improve the quality of replies from the following aspects: one is to improve the retrieval matching success rate,the second is to make the generated replies with background knowledge information,and the third is to fine-grained sense of context.The main work of this paper is as follows:(1)A hybrid model ultilizing retrieval-based chat and generative-based chat is proposed in this thesis.In order to improve the quality of replies obtained,this paper proposes a hybrid method of retrieval and generation.The existing conversation system is generally a retrieval conversation or a generative conversation.It only uses the retrieval conversation.Although the quality of the reply statements is high,it is limited to the size of the corpus and cannot "creatively" generate the reply;While only using generative conversation can generate replies "creatively",the quality of replies generated is low.For this reason,this paper designs a hybrid model of retrieval and generation,which mixes the statements of the output of the retrieval structure and the statements of the output of the generation structure,and finally makes a sort selection.The hybrid model can effectively combine the advantages of the two methods.(2)Propose a situational information fusion strategy,using the contextual state information inherent in the sentence itself in the retrieval model.In the conversation model,the situation state information contained in the sentence in the conversation context is crucial for matching the reply.Most of the actual conversations are semantically consistent and also have state consistency.The text can be more accurately retrieved and matched by incorporating the situational status information in the sentence,and the correct reply can be more easily retrieved.This paper has carried out comparative experiments on session data sets Douban,E-Commerce,and Ubuntu.The results show that the model proposed in this paper can obtain higher retrieval accuracy and more appropriate responses than the baseline model.(3)Propose the knowledge fusion strategy,integrate external common sense knowledge into the generative model.Because the quality of replies generated by the existing model is generally uncontrollable,the sentences contain less information,the semantic richness is low,and there is insufficient common sense knowledge,this paper introduces the common sense knowledge map to obtain reply statements with richer semantics,and also obtains better replies than the baseline model in the above three data sets,with higher semantic richness.(4)Based on the session hybrid model proposed in this paper,a session generation system is designed and implemented,which can generate appropriate reply content for the session content. |
