Research On Conversation In Combination Of Retrieval And Generation

Dialogues are divided into single rounds and multiple rounds in the category,usually in a question-and-answer manner.There are three ways to implement the dialogue model: the first is a dialog model with rules matching,that is,a dialogue model that relies on manually formulated rules to respond to the problem,and the second is a dialogue model based on retrieval,that is,through similarity calculation.The correct response is found,and the third is based on the generated dialog model,which is a dialog model that automatically generates responses based on the given context.At present,the main conversation models used are the retrieval dialogue model and the production dialogue model.The advantages of the retrieval dialogue model and the generic dialogue model are reflected in different aspects.Therefore,this paper hopes to develop a retrieval dialogue model that can better exert their respective advantages.And the generative dialogue model,at the same time find the appropriate combination,combine the advantages of the two dialogue models to ultimately improve the performance of the dialogue model,and studied in the following three directions:In this paper,the dialogue model based on retrieval method is studied firstly.In the aspect of Baseline,the traditional retrieval model based on TF-IDF and the retrieval dialogue model using cyclic neural network are studied,that is,the representation of sentences is obtained through the cyclic neural network,and then used.The method of similarity calculation such as dot product obtains the similarity between the problem and the candidate reply to select the reply with the highest score.Finally,the effect of using the self-attention mechanism to encode the sentence to obtain a deeper vector representation on the effect of the dialogue model is studied.After that,this paper studies the dialog model based on the generation method.In the aspect of Baseline,the sequence-to-sequence generation dialogue model is studied.The LSTM and BiLSTM networks are used as the codec model structure.Then,the attentiongrabbing mechanism is studied.The sequence generation dialogue model,this paper studies the influence of external knowledge on the model generation results,respectively studies the influence of external information on the model results and the use of situation information and the combination of the two on the model results,and studies the training techniques such as Teacher Force.How to improve the effect of the model,how some prediction techniques affect the experimental results.Finally,this paper studies the combination of retrieval and generation.In the combination,this paper studies two different ways of combining retrieval and generation: the first is to set the threshold and use the threshold as the boundary value to judge the retrieval.The answer score and the threshold value determine the combination of which answers to use.The second is to fuse the information generated by the model into the retrieval model.How to integrate the information of the retrieval model into the generation model to enhance the effect of the model,and compare the results of the experiment.The effect of the two fusion methods.