Multi-turn Response Selection Of Dialogue Chat System Based On Hierarchical Residual Matching Network

In recent years,thanks to the rapid development of artificial intelligence,great breakthroughs have been made in the construction of dialogue chat systems.The existing methods for building dialogue and chat systems can be roughly divided into two categories:generation-based methods and retrieval-based methods.The generation-based methods can generate highly consistent new responses in a given session environment.The retrieval-based methods try to find the most relevant context-response pairs in the pre-constructed dialogue corpus when given some conversation contexts.In order to improve the response accuracy of retrieval-based dialog chat system,this paper improves the existing sequence modeling method and proposes a Hierarchical Residual Matching Network(HRMN).Whether it is a generation-based dialogue chat system or a retrieval-based dialogue chat system,it is necessary to conduct sequence modeling for the conversational.In sequence modeling,making full use of memory is one of the challenges in constructing memory-based models..Existing work either as recurrent neural network,memory capacity is too small to comprehensively model information of the sequence,or as the memory network,although with an external storage structure to enhance memory,the memory is not sufficiently utilized.To address these issues,we propose a novel Memory Interactive Recurrent Unit,which constructs a multi-dimensional memory inside the recurrent unit and employs convolution operations to interact and update memories.Therefore,it can distill more information from the sequence and take full advantage of memory.Response selection in retrieval-based chatbot aims to find the most relevant response in a candidate repository given conversation context.A key technique to this task lies in how to measure the matching degree between conversation context and response at rich semantic information.In this issue,we propose a Hierarchical Residual Matching Network to make full use of the rich semantic information in the conversation history and response for multi-turn response selection task.Firstly,Hierarchical Residual Matching Network distills variously hierarchical semantic information through the well-designed multiple semantic encoder layer.Secondly,it is a deep architecture,which gets stuck in training.To alleviate this issue,we explore and consider two kinds of residual techniques.Thirdly,we adopt bi-directional attention flow to collect the matching features at all hierarchical and residual semantic information.Finally,the matching features will be separately to calculate the sub-matching score and the last match score is the sum of them.We empirically verify Hierarchical Residual Matching Network on two benchmark data sets and compare against advanced approaches.Evaluation results demonstrate that it has a steep improvement in response selection.