Research On Retrieval Dialogue Model Based On Deep Learning

The dialogue system is a software system that can conduct interactive conversations with humans in natural language.The dialogue system has been widely used in the fields of company customer service,medical and health,etc.,bringing huge commercial value to mankind.Dialogue system response models are usually divided into two categories: retrieval-based dialogue models and generation-based dialogue models.The retrieval-based dialogue model is mainly based on the dialogue context to select a response that best matches the content of the current dialogue from the candidate response set,which is divided into single-turn response selection and multi-turn response selection.Single-turn response selection only selects the response based on the current message,and multi-turn response selection based on the context of the dialog.In the selection of multi-turn response selection based on retrieval,the current message of the dialogue(the content of the last turn of dialogue in the context)and the content related to it in the dialogue context are critical to the choice of response.At the same time,there is still a problem that the reply information has nothing to do with the interlocutor’s own situation and the amount of information is small.Therefore,this paper studies the dialogue system based on additional information and adds persona information to the retrieval dialogue model.However,the existing research methods based on persona information are insufficient in extracting the semantic information of the persona information,and the accuracy rate is low.In view of the above-mentioned problems in the research of multi-turn response selection in retrieval-based conversation,this paper conducts research from the following two aspects:(1)This paper proposes a context-aware retrieval dialogue model,which learns the feature vector representation of dialogue context and response,and performs reply matching.First,the model use Bi LSTM to encode the dialogue context and candidate responses.Secondly,the attention mechanism is used to extract the important features of each sentence in the context.Then separate the current message from the dialogue context(the last utterance in the dialogue context),and calculate the degree of relevance between the current message and each utterance in the context to obtain the aggregated representation of the utterance.Finally,MLP is used to predict the score of each candidate response.Experiments on two English public data sets and one Chinese data set show that the performance of the model is better than the baseline model in most indicators,and demonstrates the cross-language compatibility of the model.(2)This paper proposes a persona-aware retrieval dialogue model based on Transformer.The model uses the encoder in Transformer to encode context,candidate responses and personal information to extract richer features.The model adopts a dual matching architecture,which performs interactive matching between responses and contexts and between responses and personas respectively for ranking response candidates.The personal information is processed by an attention mechanism based on the current message,content related to it in the context,and candidate responses to extract more effective personal feature.The comparative experiment results on the PERSONA-CHAT dataset show that the method proposed in this paper surpasses a large number of baseline models in most metrics.It can make full use of personal information to enhance the performance of model response selection.