Research About Question Answering Based On Deep Learning

With the development of computer technology and the prosperity of the Internet,information resources are becoming increasingly richer and richer.A large amount of text data exist on the Internet in the form of text,which bring great convenience to people's life.How to quickly obtain valuable information has become a difficult problem.Traditional question answering systems can help users to obtain information by using the existing knowledge storage,but it cannot accurately locate the user's intention to achieve rapid and efficient interactive requirements.By contrast,intelligent question answering systems can mine the potential semantic information in the form of natural language,and quickly and efficiently meet users' information needs.In this paper,we study Chinese answer selection task.The task description is listed as follows:given a question q and its candidate answers,the goal is to find the answers that best match the question q from the candidate answers.In this task,the length of questions and answers usually is not fixed,and a question can have more than one correct answer.The nature of this task is to calculate the semantic similarity between the question and the answer.In recent years,deep learning has been widely used in the field of natural language processing.Compared with traditional machine learning,it does not rely on manual extraction of features,language tools or additional knowledge.Therefore,we studies question answering technology based on deep learning.In our work,the basic answer selection model is:BiLSTM model based on attention mechanism.On the basis of this model,we have improved from the two perspectives of embedding representation and model extension.Firstly,for embedding representation,we proposed a word and character mixed embedding approach for generating richer vector representations of questions and answers.Secondly,for model extension,we have actively explored from the aspects of deep learning model combination and introduction of additional knowledge:the former further enriches the feature representation of questions and answers by combining BiLSTM and CNN models;the latter supplements the topic feature for deep learning model by adopting mixed attention mechanism.The experimental results on the 2016 NLPCC QA test set show that the improved answer selection models out-performs the basic model.Among them,the optimal composite model reached 79.63%,79.56%,and 69.60%on the evaluation indicators MRR,MAP,and ACC@1,respectively.