Computational Aspects of Discourse Analysis

动机“写一份NSERC DG申请需要时间……简在两个月内写好了她的申请。”在连贯的语篇中，语篇单位不是孤立地理解的，而是通过话语关系相互联系的，这些关系可能被明确标记，也可能不被明确标记。Jane在两个月内编写了她的应用程序，这一事实“说明”编写NSERC DG应用程序是很长的！语篇分析的研究试图建立这种关系的模型，这使得我们能够解释文本并理解其单位的交际目的。这反过来又对许多自然语言处理应用程序有用，例如自动摘要、文本简化...长期的客观语篇连贯指的是语篇单位之间的逻辑联系(如“例证”、“结果”、“目的”……)。尽管近年来已经做了很多工作，但计算语篇分析仍处于起步阶段，许多重要的悬而未决的问题仍然存在。这项研究计划的长期目标是探索自动语篇分析的计算方面。科学方法本研究将探讨三个具体问题：1)体裁对语篇标记的影响：在这里，我们将探索语篇体裁(例如，政治的、程序的、科学的，……)影响话语关系的使用。在需要捕获和建模的领域和流派之间的关系的语言实现方面存在显著差异。我们将讨论的问题类型包括：语篇体裁对关系的使用和话语标记语的选择有什么影响？全球话语结构与地方关系之间的互动是什么？我们能确定当地关系的刻板印象模式，特别是文本类型吗？这项工作将确定和衡量语言特征(词汇、语境、句法和语义信息)和语篇体裁之间的相关性，这些语篇体裁将被用来为特定体裁定制语篇标记。2)跨语言的无监督语篇标注：这里我们将调查关系的用法和语言实现在不同语言之间是如何变化的。我们将对英语和法语中的语篇关系进行跨语言对比，并重点讨论两个问题：语篇关系的使用是否独立于语言？跨语言的语篇关系是如何对齐的？这项工作将产生一个大型的跨语言资源，我们将向研究界提供这些资源，并将用于从另一种语言的其他分析器归纳出一种语言的语篇分析器。3)使用语篇分析进行文本简化：在这里，我们将探索如何使用语篇分析来自动改进简化文本，使其能够被更广泛的受众所理解，而不管他们的语言技能如何。特别是，我们将解决两个问题：我们能否通过更明确地表示隐含关系来实现文本简化？我们可以通过修剪信息量较小的文本跨度来实现文本简化吗？这项工作的意义随着自然语言处理研究的现状，语篇分析的计算方面的工作现在已经成为可能，并引起了研究界的极大兴趣。解决这一问题将使我们能够超越理解文本的字面意义，而是解读其“深层”意义或交际意图。具体地说，我们的工作可以有益于加拿大语言行业的所有部门：在语音处理、机器辅助翻译、内容管理、语言电子学习……有了所有这些潜在的应用，我相信我们的工作可以很容易地转移到加拿大的语言行业。

MOTIVATION "Writing an NSERC DG application takes time... Jane wrote hers in two months." In a coherent text, textual units are not understood in isolation but in relation with each other through discourse relations that may or may not be explicitly marked. The fact that Jane wrote her application in two months, "illustrates" that writing a NSERC DG application is long! Research on discourse analysis tries to model such relations, which allows us to interpret the text and understand the communicative purpose of its units. This, in turn, is useful for many Natural Language Processing applications such as automatic summarization, text simplification... LONG TERM OBJECTIVE Discourse coherence refers to the logical connections between textual units (for example, "illustration", "result", "purpose"...). Although, much work has been done in recent years, computational discourse analysis is still in its infancy and many important open questions still remain. The long term objective of this research program is to explore computational aspects of automatic discourse analysis. SCIENTIFIC APPROACH This research will explore 3 specific questions: 1) The exploration of the effect of genre on discourse tagging: Here, we will explore how the textual genre (e.g. political, procedural, scientific,...) affects the use of discourse relations. There are significant differences in the linguistic realization of relations across domains and genres that need to be captured and modeled. The types of questions we will address include: What is the influence of the textual genre on the usage of relations and choice of discourse markers? What is the interaction between global discourse structures and local relations? Can we identify stereotypical patterns of local relations in particular textual genres? This work will identify and measure the correlation between linguistic features (lexical, contextual, syntactic and semantic information) and textual genres which will be used to tailor discourse taggers to a specific genre. 2) Unsupervised discourse tagging across languages: Here we will investigate how the usage and linguistic realization of relations vary across languages. We will perform a cross-lingual comparison of discourse relations in English and French, and focus on two questions: Is the usage of discourse relations language-independent? How do discourse relations align across languages? This work will result in a large cross-lingual resource which we will make available to the research community and will be used to induce discourse parsers in one language from other parsers in another language. 3) The use of discourse analysis for text simplification: Here we will explore how discourse analysis can be used to improve simplify texts automatically to make them accessible to a wider audience regardless of their language skills. In particular, we will address two questions: Can we achieve text simplification by signaling implicit relations more explicitly? Can we achieve text simplification by pruning less informative text spans? SIGNIFICANCE OF THE WORK With the current state of the art in Natural Language Processing research, work on computational aspects of discourse analysis has now become possible and is growing much interest in the research community. Addressing this issue will allow us to go beyond understanding the literal meaning of a text, but interpret its "deep" meaning or communicative intention. Concretely, our work can be beneficial to all sectors of the Canadian language industry: in speech processing, machine-aided translation, content management, language e-learning... With all these potential applications, I am confident that our work can easily be transferred to the Canadian language industry.