Study On Multi-turn Response Selection For Retrieval Chatbots Based On Deep Learning

With the continuous development of online shopping,after-sales customer service,educational consulting and other fields,traditional service methods based on manual customer service have shown disadvantages such as high labor cost and uneven customer service quality.Meanwhile,with the significant improvement of data accumulation and computing power,chatbots based on deep learning have shown an explosive growth trend,showing huge economic advantages over human customer service.At present,the research of chatbot has been paid much attention and become a research hotspot.Multi-turn response selection is the key to implementing retrieval chatbot.There are still some problems in current research,such as the lack of mining word sequence information and the lack of considering that the importance of different text information in the context is different.To solve the above problems,this thesis makes an in-depth study on the multi-turn response selection model.The main research work is as follows:(1)A multi-turn response selection model based on multi-level word sequence granularity representation and fusion word vector(Sequential Matching Network with Multi-Level Granularity Representations,MRSMN)is proposed.In this model,multi-level word sequence granularity is used to solve the problem of insufficient word sequence information mining,and fusion word vector is used to solve the problem of information loss and overfitting of original word vector caused by only using training word vector in the existing model.The experimental results show that the evaluation metrics of MRSMN have been improved on the two experimental datasets.(2)A multi-turn response selection model based on bidirectional attention and spatio-temporal matching feature(Multi-Level Granularity Representations with Bidirectional Attention and Spatio-Temporal Matching Feature,MRBAST)is proposed.This model uses bidirectional attention to extract the information that different text content in the context has different importance for the response and measure the relevance between the candidate response and the overall semantics of the context.On the other hand,the model uses the spatio-temporal matching feature extracted by 3D convolutional neural network to solve the problem that the model cannot be trained after adding attention representation.Experiments show that the performance of MRBAST model is better than MRSMN and better than most benchmark models.In this thesis,the validity of the proposed MRSMN model based on multi-level word sequence granularity representation and fusion word vector and the MRBAST model based on bidirectional attention and spatio-temporal matching feature is verified by experiments on the E-commerce dialogue corpus and Douban dialogue corpus.