Research On Retrieval-based Dialogue System Based On Knowledge Fusion Matching Network

In recent years,knowledge-grounded dialogue systems in the open domain have attracted the attention of researchers,whose goal is to use the background knowledge of the conversation and the conversation context to answer users'questions.The knowledge-based dialogue systems can be categorized into two groups:generative one and retrieval-based one.The former,which is based on encoder-decoder framework,integrates background knowledge and then generates answers.The latter is to select the response that best matches the background knowledge and the conversation context from the candidate responses set.In this paper,we focus on knowledge-grounded retrieval-based dialogue systems.The existing methods focus on using a dual interactive matching framework,which performs interactive matching between responses and contexts and between responses and background knowledge respectively.Two matching features are combined in the final stage.However,these methods ignore the effect of blackground knowledge in the response-context matching stage.The two separate matching stages limit the recognition of semantic features of the model,which may limit the improvement of the model performance.To solve these problems,this paper proposes a retrieval-based dialogue system model based on knowledge fusion matching network(KFMN).The model first constructs a knowledge selection and fusion module,selects the relevant background knowledge for each utterance in the context and response candidate,and uses the attention mechanism to integrate the relevant background knowledge into each utterance and response candidate.The operation enhances the semantic representation of each sentence.At the same time,related dialogue content information is integrated into each background knowledge sentence to achieve bidirectional semantic enhancement.In the matching stage,response candidate matches with the conversation context and background knowledge respectively.Since the existing methods only use the matching features of sentence dimension in the last stage and ignore the matching features of other dimensions,KFMN further extracts the matching features of word dimension from the response-context matching matrix to provide more matching signals for the model.The experimental results show that KFMN can effectively improve the effect of knowledge-grouneded retrieval-based dialogue systems.On the original and the revised version of the Persona-Chat dataset,KFMN achieves 80.3%and 73.0%on₂₀@1,respectively;On the CMUDo G dataset,KFMN achieves 80.9%on₂₀@1.