Metagrammatical Knowledge for Grammars and Corpora

关于我们在自然语言中遇到的句法现象，目前在理论语言学家(各种框架)和自然语言处理(NLP)研究人员中已经达成了广泛的共识。然而，有许多不同的框架对这些现象进行了分析，甚至在一个框架内进行具体分析也存在分歧。因此，语言资源(如带注释的语料库或语法)不能跨框架轻松重用。这个项目将调查语言学和自然语言处理工作中常见的句法分类。这种基本的分类被称为“元语法”。在给定元语法的情况下，可以生成一个工具来在不同的框架中自动生成语法。本研究包括三个主要活动。第一个涉及几种语言(包括英语)的比较工作，这将导致这些语言的元语法协调一致。这些独立于框架的规范将对句法属性进行分类，并详细说明它们可能的交互作用；不同语言之间共享的类别将导致元语法的共享部分。第二个是关于特定语法陈述的发展，这些语法陈述将元语法范畴与特定框架和特定语言中的构式联系起来。正是这些陈述，在它们的相互作用中，决定了词序。第三个项目是使用元语法中的句法特征对宾夕法尼亚树库语料库进行标注，从而使隐含在宾夕法尼亚树库短语结构中的信息显性地被其他框架所使用。这个项目将使自然语言规划和语言学界更好地分享对句法现象的见解。此外，这项工作将使开发新的NLP工具变得不那么依赖于特定的表示。它将使语言学家能够为不同的框架和语言迅速开发语法和测试套件，从而允许对语言语法进行跨框架和跨框架的评估。在该项目完成后，将向研究界提供重新注释了元语法高级别类别的PTB。

There is today a broad consensus among theoretical linguists (of all frameworks) and researchers in Natural Language Processing (NLP) about what the syntactic phenomena are that we encounter in natural languages. However, there are many different frameworks in which analyses of these phenomena have been implemented, and there is even disagreement about specific analyses within one single framework. As a result, linguistic resources such as annotated corpora or grammars cannot be easily reused across frameworks. This project will investigate the common categorization of syntax that underlies work in linguistics and NLP. This underlying categorization is called a ``metagrammar''. Given a metagrammar, a tool can be produced to automatically generate grammars in different frameworks. This research contains three main activities. The first involves comparative work in several languages (including English) that will lead to coordinated metagrammars for these languages. These framework-independent specifications will catalog syntactic properties and detail their possible interaction; categories shared between languages will lead to shared portions of the metagrammar. The second concerns the development of specific grammar statements that relate metagrammatical categories to constructs in particular frameworks and for particular languages. It is these statements that, in their interaction, determine word order. The third involves annotating the Penn Treebank (PTB) corpus with the syntactic properties from the metagrammar, thus making the information implicitly encoded in the phrase structure of the PTB explicit and usable by other frameworks.This project will enable the NLP and linguistics communities to better share insights on syntactic phenomena. Additionally, the work will enable the development of new NLP tools that are less dependent on a particular representation. It will enable linguists to rapidly develop grammars and test-suites for different frameworks and languages, thus allowing for both cross- and inter-framework evaluation of linguistic grammars. Upon completion of the project, the PTB re-annotated with the high-level categories of the metagrammar will be made available to the research community .