Research On Particular-form Oriented Machine Reading Comprehension

Machine reading comprehension is a hot research task in current natural language processing which aims at enabling machines understand text document and further answer the related questions correctly.The improved ability to understand text for machines will benefit information retrieval,question answering and machine translation a lot and,meanwhile,will do good to get a better user experience on search engine and intelligent assistant.To conclude,carrying out researches about text understanding on machine reading comprehension task is valuable both in research and application.There are various forms of machine reading comprehension.We choose spanbased and multi-choice machine reading comprehension to carry out our researches.With the rapid progress on computation power and availability of large-scale dataset,we thus construct models based on deep learning and reinforcement learning to learn abstractive representation of document,question and candidate answers.Then develop and evaluate machine comprehension techniques.Plenty of deep learning based machine reading comprehension papers have been published.We also learn representation of document and question with the help of deep learning.And try to empower machines to understand information in document and question via operations on representations like multi-turn interaction,gated information filtering,context matching.We simultaneously combine various attention mechanisms into our model to merge information and predict answers.Deep reinforcement learning is regarded as the future of deep learning and the vital means towards AI-complete.We apply DQN algorithm,one of deep reinforcement learning algorithm,to span-based machine reading comprehension.The design of actions,environment state representation,reward policy and QNetwork architecture has been done to transform machine reading comprehension.As a discourse-level natural language processing task,the document in machine comprehension task contains discourse-level semantics and more information comparing to sentence-level tasks.And reasoning ability is crucial to understand document directed by question through synthesizing diverse information and discourse semantics.We apply context similarity matching deep learning model to multi-choice machine reading comprehension.Analysis and evaluate the model’s reasoning ability.