Research On Question Answering Technology In Text Reading Tasks Based On Neural Networks

Making machines understand human language and perform interactive question answering with natural language has always been a pursuit of human being.Question Answering(QA)for text reading tasks,one critical technique under this challenge,has always been one of the main research areas in natural language processing.Given one passage,the algorithm is required to answer the question automatically based on its understanding of the question.To answer it correctly,certain background knowledge,thorough understanding of the question,and appropriate answer generating method are needed.With the development of deep learning technology,Neural Networks(NN)have gain widely attention in speech signal processing and natural language processing.Ques-tion answering techniques based on Recurrent Neural Networks(RNN)have lots of benefits.They can memorize and understand longer text,effectively search the content with the given question and generate the answer,therefore largely improve the accuracy and outperform the time-consuming traditional rule-based methods.In this dissertation,we focus on the research of neural network-based question an-swering technique in text reading tasks.We choose the Stanford Question Answering Dataset(SQuAD),one of the most impactful datasets in question answering area,as the dataset for experiments.Firstly,we propose a conditional encoder-decoder frame-work based on recurrent neural networks and apply this framework to SQuAD,which achieves superior performance compared with the official baseline using traditional rule-based method.Based on this conditional encoder-decoder framework,we further study the problem of question understanding,and propose a new question understand-ing framework based on TreeLSTM and word frequency filter banks.Better question understanding helps the algorithm find the right answer better and hence improves the accuracy,as proved by the experiments.Furthermore,most existing models are trained for all question types,while questions of different types have their own features that differ from the others.To model different question types and find the pattern of correct answer,we propose an end-to-end trainable solution for question adaptation.Our ap-proach combining these three research ranks No.2 in the SQuAD official hidden dataset.