Research Of Machine Reading Comprehension For Open-Domain Question Answering

Nowadays,information technology is constantly undergoing great innovation.Data resources are also showing an explosive growth trend,simultaneously.Gradually it has become a hot topic concerned by academia and industry that how to obtain information efficiently and accurately.The initial information retrieval and analysis tasks are completed by search engines,which retrieve a large number of web pages on the internet based on keywords and return the most relevant set of web pages to users.Regrettably,such search results are still too large in scale to allow users to quickly find accurate answers.Therefore,how to efficiently obtain accurate answers from the internet becomes a key issue leaving to be solved urgently.Accurate answers can be generated from massive data by open domain question answering technology based on machine reading comprehension.This technology adopts the "retrieve-read interactive framework" to obtain relevant documents through the retriever,to understand relevant documents based on the reader,and to adjust the retrieval direction according to the results of reader until the final answer is obtained.Most of the existing reading comprehension algorithms are proposed under supervision,ignoring a large amount of unlabeled data in actual scenes.In addition,the existing interactive framework fixed the number of interactive rounds of retrieval and reading,and still introduced the noise document after finding the correct answer,which reduced the efficiency and accuracy of QA.This thesis proposed a novel open domain question answering algorithm based on reading comprehension,which optimized the retriever and the reader by alleviating the above problems.For the extractor,this thesis proposes a multi-passage machine reading comprehension algorithm based on multi-task learning and generative adversarial training.In this algorithm,in order to enable the generative adversarial training process to be trained end-to-end,a hybrid answer extraction method is proposed to produce the answer candidates set and its representation.In order to verify the accuracy of the candidate answers using the original passages,a global-local memory augment neural network is proposed to encode the passage set.The experimental results on multi-passage machine reading comprehension datasets confirm the effectiveness of the proposed algorithm.For the retriever,this thesis proposes an interactive framework based on reinforcement learning.In this algorithm,an interactive control module based on reinforcement learning is proposed to dynamically control the number of interactive rounds.In order to improve the overall quality of the retrieved passage set,this thesis proposed a global optimization function of passage set based on policy gradient.The experimental results on question answering dataset show that the proposed algorithm improves the efficiency and accuracy significantly compared with the existing work,verifying the effectiveness of the proposed algorithm.