Research On Machine Reading Comprehension Using Multiple Semantic Alignment Representations

Machine Reading Comprehension(MRC)is to let the machine learn to read comprehension,that reading the text and understanding the main point of the text,and then answering the questions.MRC is an important mark of the development level of natural language processing(NLP)and has a wide range of application scenarios,including search engines,machine scoring,question answering systems,information retrieval,etc.It can enable computer to help human to find the answers in the large amount of text data,which reduce the cost of people's access to information.In recent years,MRC is a very hot topic in the field of artificial intelligence.In particular,MRC based on deep learning has been widely concerned by researchers both at home and abroad.Compared with the traditional method,the MRC with deep learning method has made significant progress.Therefore,this thesis used deep learning method and combined the characteristics of MRC tasks to implement and improve MRC system that fuses multiple semantic alignment representations.Specifically,the work of this thesis mainly includes the following three aspects:(1)This thesis detailed MRC datasets's development,common evaluation indicators,and various MRC models.this thesis summarized the existing MRC research methods into two different types:traditional methods and deep learning methods by in-depth comparison and analysis of various MRC models.(2)This thesis designed and implemented a MRC baseline system based on neural attention.Baseline system uses a deep learning method and attention mechanism to implement an end-to-end neural network model.The baseline system first uses convolutional neural networks and pre-training method to obtain char-level embedding and word embedding,then imploits bidirectional long short-term memory model to encode the context of passage and question,next uses attention mechanisms to obtain questionto passage's semantic alignment representation,finally model's output layer uses softmax classifier to caculate answer's probability distribution.(3)This thesis presented a MRC model using multiple semantic alignment representations.The system mainly improves baseline model,which enhances semantic expression and gets better interactive representation of passages and questions by exploiting multiple semantic alignment representations.The system makes three improvements to the baseline system.First,this system introduce an enhanced factor to enhance passage's semantic representation,whch can obtain semantic alignment representation of from passage to the question direction.Second,this system increase the self-alignment semantic representation of from passage to passage directionn?Third,output layer reintroduces the question information and uses the attention to connect the start and end of the answer.The experimental results show that the performance of the improved model is obviously better than that of baseline model.