Rediscovering Annotation Projection for Cross-Lingual Parser Induction

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1854–1864, Dublin, Ireland, August 23-29 2014. Rediscovering Annotation Projection for Cross-Lingual Parser Induction J¨org Tiedemann Uppsala University Department of Linguistics and Philology firstname.lastname@lingfil.uu.se Abstract Previous research on annotation projection for parser induction across languages showed only limited success and often required substantial language-specific post-processing to fix inconsis- tencies and to lift the performance onto a useful level. Model transfer was introduced as another quite successful alternative and much research has been devoted to this paradigm recently. In this paper, we revisit annotation projection and show that the previously reported results are mainly spoiled by the flaws of evaluation with incompatible annotation schemes. Lexicalized parsers created on projected data are especially harmed by such discrepancies. However, recently de- veloped cross-lingually harmonized annotation schemes remove this obstacle and restore the abilities of syntactic annotation projection. We demonstrate this by applying projection strate- gies to a number of European languages and a selection of human and machine-translated data. Our results outperform the simple direct transfer approach by a large margin and also pave the road to cross-lingual parsing without gold POS labels. 1 Introduction Linguistic resources and tools exist only for a minority of the world’s languages. However, many NLP applications require robust tools and the development of language-specific resources is expensive and time consuming. Many of the common tools are based on data-driven techniques and they often re- quire strong supervision to achieve reasonable results for real world applications. Fully unsupervised techniques are not a good alternative yet for tasks like data-driven syntactic parsing and, therefore, cross- lingual learning has been proposed as a possible solution to quickly create initial tools for otherwise unsupported languages (Ganchev and Das, 2013). In syntactic parsing, two main strategies have been explored in cross-lingual learning: annotation pro- jection and model transfer. The first strategy relies on parallel corpora and automatic word alignment that make it possible to map linguistics annotation from a source language to a new target language (Yarowsky et al., 2001; Hwa et al., 2005; T¨ackstr¨om et al., 2013a). The basic idea is that existing tools and models are used to process the source side of a parallel corpus and that projection heuristics guided by alignment can be used to transfer the automatic annotation to the target language text. Using the projected annotation assuming that it is sufficiently correct, models can then be trained for the target language. However, directly projecting syntactic structure results in a rather poor performance when applied to resources that were developed separately for individual languages (Hwa et al., 2005). Exten- sive additional post-processing in form of transformation rules is required to achieve reasonable scores. Furthermore, incompatible tagsets make it impossible to directly transfer labeled annotation to a new language and previous literature on cross-lingual parsing via annotation projection is, therefore, bound to the evaluation of unlabeled attachment scores (UAS). Less frequent, but also possible, is the scenario where the source side of the corpus contains …

Content truncated to 500 words