基于深度神经网络的端到端语篇解析

发布时间：2018-07-31 21:04

【摘要】：语篇关系是指两个文本单元之间的语义逻辑关系,如因果关系、转折关系等。语篇解析以自由文本为输入,返回文本中蕴含的语篇关系。语篇解析可以广泛应用于许多自然语言处理任务,如问答系统、统计机器翻译和自动文摘等。以往的研究工作多集中在语篇解析的各个子任务,如语篇连接词识别、论元标记、显性或隐性语篇关系语义识别等。本文的第一个工作是构建一个精心设计的英文语篇解析器,在Lin等[7]的语篇解析器基础上,提出并设计更多有效特征和组件,这一系统在2015年的CoNLL英文语篇解析竞赛中获得第一名。为进一步构造面向实用的显性语篇解析器,本文分别对两个论元抽取不同特征,构建不同的分类器,显著提升性能,这一部分工作在2015年的KSEM(CCF-C类)会议上获得"最佳论文奖"。由于缺乏中文标注语料,中文语篇关系的相关研究很少。随着CDTB语料库的发布,本文的第二个工作是深入分析中英文的不同特点,构建一个有效的中文语篇解析器,这一系统在2016年CoNLL的中文语篇解析竞赛中获得第二名。隐性语篇解析是整个语篇解析的瓶颈,以往采用语言学特征结合机器学习的方法,需要专家领域知识,系统泛化性差。随着深度学习在自然语言处理领域的深入发展,本文的第三个工作首先提出使用卷积神经网络进行隐性语篇关系的语义识别,该系统在2016年CoNLL的英文语篇解析竞赛中获得第二名。进而,本文又提出基于关注机制的神经网络来提升隐性语篇关系语义识别性能。深度学习需要大量的训练样本,为了解决人工标注数据量的局限性,本文的第四个工作提出多任务的神经网络模型,利用大量人工合成的隐性语篇关系数据,进一步提升识别性能。第三和第四个工作目前已投稿2017年的EMNLP会议。本文在中文和英文语篇解析上,进行了大量的实验。实验结果表明,本文提出的端到端英文和中文语篇解析的有效性,进而采用基于关注机制和多任务的深度神经网络模型,能够帮助提升中英文隐性语篇关系的语义识别性能。
[Abstract]:Textual relations refer to the semantic logical relationships between two text units, such as causality, transition, etc. Text parsing takes the free text as input and returns the textual relationship contained in the text. Discourse parsing can be widely used in many natural language processing tasks, such as question-answering systems, statistical machine translation and automatic abstracts. Previous studies have focused on the subtasks of discourse parsing, such as textual conjunction recognition, argument meta-marker, explicit or implicit textual relation semantic recognition, and so on. The first work of this paper is to construct a carefully designed English text parser, which proposes and designs more effective features and components on the basis of Lin et al. [7]. This system won the first place in the CoNLL English discourse Analysis Competition in 2015. In order to construct a pragmatic explicit discourse parser, this paper extracts different features from two arguments, constructs different classifiers, and improves performance significantly. This part of the work won the "best paper award" at the KSEM (CCF-C class) conference in 2015. Due to the lack of Chinese tagging data, there is little research on the relationship between Chinese texts. With the release of the CDTB corpus, the second work of this thesis is to analyze the different features of Chinese and English, and construct an effective Chinese text parser, which is the second place in the 2016 CoNLL Chinese discourse parsing contest. Implicit discourse parsing is the bottleneck of the whole text parsing. In the past, the method of combining linguistic features with machine learning requires expert domain knowledge, and the system generalization is poor. With the development of deep learning in the field of natural language processing, the third work of this paper first proposes the use of convolutional neural networks for semantic recognition of recessive discourse relations. The system won the second place in the 2016 CoNLL English discourse Analysis contest. Furthermore, this paper proposes a neural network based on attention mechanism to improve the semantic recognition performance of recessive discourse relations. In order to solve the limitation of artificial tagging data, the fourth work of this paper proposes a multi-task neural network model, which uses a large number of artificial synthetic recessive text relational data. Further improve the recognition performance. The third and fourth jobs are currently contributing to the 2017 EMNLP meeting. In this paper, a large number of experiments have been carried out in Chinese and English discourse analysis. The experimental results show that the proposed end-to-end English and Chinese discourse parsing is effective, and the depth neural network model based on attention mechanism and multitasking can help to improve the semantic recognition performance of implicit discourse relations in both Chinese and English.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP18

【相似文献】