||
本文为美国斯坦福大学(作者:DanqiChen)的博士论文,共156页。
教会机器理解人类语言文档是人工智能中最难以捉摸和长期存在的挑战之一。本文探讨了阅读理解的问题:如何建立计算机系统来阅读课文并回答问题。一方面,我们认为阅读理解是评估计算机系统对人类语言理解程度的一项重要任务。另一方面,如果我们能建立高性能的阅读理解系统,它们将成为问答和对话系统等应用的关键技术。
本文主要研究神经阅读理解:一类基于深度神经网络的阅读理解模型。与传统的稀疏的、手工设计的基于特征的模型相比,这类端到端的神经模型在学习丰富的语言现象和提高现代阅读理解基准的性能方面有着更大的优势。
本文由两部分组成。在第一部分中,我们旨在阐述神经阅读理解的本质,并阐述我们在构建有效的神经阅读理解模型方面所做的努力,更重要的是,了解神经阅读理解模型实际学习的内容,以及解决当前任务所需的语言理解深度。我们还总结了该领域的最新进展,讨论了未来的发展方向和开放性问题。
在本论文的第二部分中,我们研究了如何在最近取得成功的神经阅读理解的基础上构建实际应用。特别是,我们开创了两个新的研究方向:1)如何将信息检索技术与神经阅读理解相结合,以解决大规模开放领域的问题解答;2)如何从当前的单轮、基于跨度的阅读理解模型中构建会话式问题解答系统。我们在DRQA和COQA项目中实现了这些想法,并证明了这些方法的有效性。我们相信它们在未来的语言技术应用中有着很大的希望。
Teaching machines to understand humanlanguage documents is one of the most elusive and long-standing challenges inArtificial Intelligence. This thesis tackles the problem of readingcomprehension: how to build computer systems to read a passage of text andanswer comprehension questions. On the one hand, we think that readingcomprehension is an important task for evaluating how well computer systemsunderstand human language. On the other hand, if we can build high-performingreading comprehension systems, they would be a crucial technology forapplications such as question answering and dialogue systems. In this thesis,we focus on neural reading comprehension: a class of reading comprehensionmodels built on top of deep neural networks. Compared to traditional sparse,hand-designed feature-based models, these end-to-end neural models have provento be more effective in learning rich linguistic phenomena and improvedperformance on all the modern reading comprehension benchmarks by a largemargin. This thesis consists of two parts. In the first part, we aim to coverthe essence of neural reading comprehension and present our efforts at buildingeffective neural reading comprehension models, and more importantly,understanding what neural reading comprehension models have actually learned,and what depth of language understanding is needed to solve current tasks. Wealso summarize recent advances and discuss future directions and open questionsin this field. In the second part of this thesis, we investigate how we canbuild practical applications based on the recent success of neural readingcomprehension. In particular, we pioneered two new research directions: 1) howwe can combine information retrieval techniques with neural readingcomprehension to tackle large-scale open-domain question answering; and 2) howwe can build conversational question answering systems from currentsingle-turn, span-based reading comprehension models. We implemented theseideas in the DRQA and COQA projects and we demonstrate the effectiveness ofthese approaches. We believe that they hold great promise for future languagetechnologies.
下载英文原文地址:
http://page2.dfpan.com/fs/9lcjb221929146e58d2/
更多精彩文章请关注微信号:
Archiver|手机版|科学网 ( 京ICP备07017567号-12 )
GMT+8, 2024-9-23 12:30
Powered by ScienceNet.cn
Copyright © 2007- 中国科学报社