Multilingual Structural Projection across Interlinear Text

Proceedings of NAACL HLT 2007, pages 452–459, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Multilingual Structural Projection across Interlinear Text Fei Xia Department of Linguistics University of Washington Seattle, WA 98195 fxia@u.washington.edu William D. Lewis Department of Linguistics University of Washington Seattle, WA 98195 wlewis2@u.washington.edu Abstract This paper explores the potential for an- notating and enriching data for low-density languages via the alignment and projec- tion of syntactic structure from parsed data for resource-rich languages such as English. We seek to develop enriched resources for a large number of the world’s languages, most of which have no significant digital pres- ence. We do this by tapping the body of Web-based linguistic data, most of which exists in small, analyzed chunks embedded in scholarly papers, journal articles, Web pages, and other online documents. By har- vesting and enriching these data, we can provide the means for knowledge discovery across the resulting corpus that can lead to building computational resources such as grammars and transfer rules, which, in turn, can be used as bootstraps for build- ing additional tools and resources for the languages represented.1 1 Introduction Developing natural language applications is generally dependent on the availability of annotated corpora. Building annotated resources, however, is a signif- icantly time consuming process involving consider- able human effort. Although a number of projects have been undertaken to develop annotated resources for non-English languages, e.g., treebanks, the devel- opment of these resources has been no small feat, and to date have been limited to a very small number of 1We would like to thank Dan Jinguji for creating the word alignment and source dependency structure gold standards. Our thanks also go to three anonymous re- viewers for their helpful comments and suggestions. the world’s languages (e.g., Chinese, German, Ara- bic, Korean, etc.). Some notable efforts have been undertaken to develop automated means for creating annotated corpora through the projection of annota- tions (Yarowksy and Ngai, 2001; Xi and Hwa, 2005). The resulting methods, however, can only be applied to a small number of language pairs due mostly to the need for sizeable parallel corpora. Unfortunately, most languages do not have parallel corpora of suffi- cient size, making these methods inapplicable for the vast majority of the world’s languages. We describe a method for bootstrapping resource creation by tapping the wealth of multilingual data on the Web that has been created by linguists. Of particular note is the linguistic presentation format of “interlinear text”, a common format used for pre- senting language data and analysis relevant to a par- ticular argument or investigation. Since interlin- ear examples consist of orthographically or phoneti- cally encoded language data aligned with an English translation, the “database” of interlinear examples found on the Web, when taken together, constitute a significant multilingual, parallel corpus covering hun- dreds to thousands of the world’s languages. We do not propose that a database of interlin- ear text alone is sufficient to create NLP resources and tools, but rather that it may act …

Content truncated to 500 words