Research On Machine Question Answering Based On Deep Learning

Machine question answering(QA)is one of the most challenging tasks in natural language processing.Although being extremely difficult,many research efforts have been dedicated to this field,due to the huge application values.This thesis focuses on deep learning based machine QA,in which,an end-to-end intelligent open-domain QA system has been proposed with current state-of-the-art neural network models in NLP.To be specific,the main contributions of this work can be described as follows:As the very core,we use reading comprehension to be the inference module of our open-domain QA system.To make the system more precise and innovative,we detailly analyzed the attention bias problem in the BERT,a promising model achieved many new state-of-the-art performance in NLP.Based on this,we propose ForceReader,which includes multiple incentive schemes such as attention separation representation,multimodal reading comprehension,conditional background attention,and deep interactive reasoning.Besides,the shortcomings of reading comprehension in machine QA applications were also explored.Relevant solutions,such as multiple information interaction,multi-task learning by combining paragraph classification and reading comprehension were proposed.With all these improvements,our inference module can capture richer language information and being able to automatically handle multi-document QA in an end-to-end way.Based on the significantly improved inference and application module,we build an intelligent open-domain QA system,which uses Wikipedia knowledge and search engines as information sources.We add question-answer library and a deep semantic matching model for problem matching,in order to quickly respond to the existing problems.In addition,the system been equipped with distant supervision and sustainable learning,so that it can continuously evolve itself during the QA feedback process.For all the improvement in our ForceReader QA system,juxtaposed experiments were designed in our working and were verified with high-quality public data such as SQuAD.Experimental results demonstrate that our proposed schemes have a significant improvement on SQuAD data,and the model has excellent performance on Chinese data.Our open domain question answering system was verified on CuratedTREC,WebQuestions and other data,and the system has presented significant improvement compared to the baseline.In order to test the system's QA ability in different languages,we migrated it to Chinese QA.With minimal migration cost,our system also demonstrates merits in Chinese data.Finally,we explore the interpretability of neural networks.Explanations were given to the knowledge learned by the model,through attention visualization.Besides showing the efficiency of the proposed designs,we also strive to break through the black box limitation.It greatly increases the interpretability of our model in a very intuitive way.