Multi-Granularity Composite Conversation Model Based On Siamese Network

Conversation generation is an important part of natural language processing and has attracted much attention in recent years.The open domain conversation model is mainly driven by data,which can be roughly divided into retrieval models or generation models.Although remarkable progress has been achieved in recent years,it is still difficult to get responses that are grammatically and semantically appropriate.The main research work is as follows:(1)This paper focuses on the technologies of Semantic textual similarity and conversation retrieval at home and abroad,and focuses on the matching algorithm for single-round sentences.Aiming at the shortcomings of previous algorithms,this paper proposes a Multi-Granularity Matching model based on Siamese Network(MGMSN).This method considers both deep semantic similarity and shallow semantic similarity of input sentences to fully mine similar information between sentences.In addition,considering the problem of OOV(Out Of Vocabulary)in sentences,this paper considers both word and character granularity in deep semantic similarity to further learn information while alleviating the problem of OOV.Finally,we conducted comparative experiments on the classic LCQMC dataset.The experimental results confirmed the effectiveness and generalization ability of the method,and the ablation experiments also showed the important role played by each part of the model.(2)Although the response obtained by retrieval models is smooth,it is difficult for retrieval models to respond to problems that do not appear in the corpus.Therefore,the paper proposes a Multi-Granularity Composite Conversation model based on Siamese Network(MGCCSN).The purpose of the MGCCSN model is to combine the retrieval model with the generation model to make up for the shortcomings of the two models.First,the input question and retrieval contexts are put into a multi-response generation module,and a number of generation candidate responses are obtained.These generated candidate responses and retrieved candidate responses further use the candidate response ranking module to select the most suitable response to the input context through pre-screening and decision making.Through automatic evaluation and manual evaluation,the composite algorithm proposed in this paper has improved relevance and diversity of response to the input problem,helping to solve the problem and the continuation of the conversation.