Repurposing Theoretical Linguistic Data for Tool Development and Search

Repurposing Theoretical Linguistic Data for Tool Development and Search Fei Xia University of Washington Seattle, WA 98195 fxia@u.washington.edu William D. Lewis∗ Microsoft Research Redmond, WA 98052-6399 wilewis@microsoft.com Abstract For the majority of the world’s languages, the number of linguistic resources (e.g., an- notated corpora and parallel data) is very limited. Consequently, supervised methods, as well as many unsupervised methods, can- not be applied directly, leaving these lan- guages largely untouched and unnoticed. In this paper, we describe the construction of a resource that taps the large body of linguisti- cally analyzed language data that has made its way to the Web, and propose using this resource to bootstrap NLP tool development. 1 Introduction Until fairly recently, most NLP research has focused on the ten or so majority languages of the world, the canonical high density languages. Low density, or resource poor languages (RPLs), have more recently captured the interest of NLP research, mostly be- cause of recent advances in computational technolo- gies and computing power. As indicated by their name, RPLs suffer from a lack of resources, namely data. Supervised learning techniques generally re- quire large amounts of annotated data, something that is nonexistent or scare for most RPLs. A greater number of RPLs, however, have raw data that is available, and the amount and availability of this raw data is increasing every day as more of it makes its way to the Web. Likewise, advances in un- and semi- supervised learning techniques have made raw data more readily viable for tool development. Still, how- ever, such techniques often require “seeds”, or “pro- totypes” (c.f., (Haghighi and Klein, 2006)) which are used to prune search spaces or direct learners. An important question is how to create such seeds for the hundreds to thousands of RPLs. We describe the construction of a resource that taps the large body of linguistically analyzed language data that has made its way to the Web, and propose using this ∗The work described in this document was done while Lewis was faculty at the University of Washington. resource as a means to bootstrap NLP tool devel- opment. Interlinear Glossed Text, or IGT, a semi- structured data type quite common to the field of linguistics, is used to present data and analysis for a language and is generally embedded in scholarly lin- guistic documents as part of a larger analysis. IGT’s unique structure — effectively each instance consists of a bitext between English and some target language — can be easily enriched through alignment and pro- jection (e.g., (Yarowsky and Ngai, 2001), (Hwa et al., 2002)). The reader will note that the IGT instance in Example (1) consists of a bitext between some tar- get language on the first line, or the target line (in this case in Welsh), and a third line in English, the translation line. The canonical IGT form, which this example is representative of, has intervening linguis- tic annotations and glosses on a second line, the gloss line. Because the gloss line aligns …

Content truncated to 500 words