Research On Neural Conversation Model Based On User Profile

At present,more and more researchers focus on dialogue systems and many companies are trying to build a dialogue system that meets their needs and business scenarios.The dialogue system has a wide range of applications in the fields of e-commerce,smart homes and so on.It can understand people's semantics and combine existing electronic devices,services to enrich people's lives.What's more for the user profile,users generate more and more information on the Internet and we can extract many user attributes and labels among it.This information has great commercial value and also has an important role in improving products.This dissertation focuses on two kinds of dialogue systems.One is question-answering system based on similarity matching models and the other is chatbot system based on generative models.We also combine the user-attribute classification model with the chatbot system to improve the pertinence and diversity of the dialogue.For the question-answering dialogue,this dissertation proposes a two-level model based on the database of frequency questions.The first level is the retrieval model which provides a reliable high-quality candidate set for the next level model.It can improve the speed of the system.We combine it with synonyms to increase the number of questions retrieved by the model.The second level is the ranking model.The dissertation proposes an attention network model which can combine multiple similarity features,different embedding representations of words,and the extended keyword information of the questions.We test the model on the artificially constructed medical question-answering dataset,and compare it with the traditional logistic regression model,the learning-to-rank model,and several common deep learning models.Both the accuracy and the MAP value of the output answers are improved.This dissertation also proposes a semi-supervised method to collect the standard questions,which can save the handcrafted costs.For the classification of user profile,we propose an attention neural network model to complete the classification of users' attributes,without relying on handcrafted features.For those user attributes related to social relationships,this model combines textual information with social relationship information to improve the accuracy.In the evaluation dataset of SMP CUP 2016,the model exceeds the state-of-art model in three tasks,including gender,age and location classification tasks.For the chatbot,this dissertation proposes a neural generative model of Encoder-Decoder structure.Both the encoding and decoding are LSTM structures and combined with the attention mechanism.In the process of generating responses,beam search is used to improve the quality of the dialogue.At the same time,we add the geographical distribution vectors generated by the user attribute classification model to the encoding and decoding process,which enables the dialogue model to learn the user's geographical information,and generates more regional features based on different geographical distribution vectors.This model has been trained on 4 million microblog comment corpus.We compare the model with other neural conversation models and provide some examples of its responses to prove the model's good performance.