What Can We Get From 1000 Tokens? A Case Study of Multilingual POS Tagging For Resource-Poor Languages

Correction: We unintentionally mis-represented Garrette et al. (2013) in the published version of this paper by stat- ing that they required an external tag dictionary. We have corrected these inaccuracies to reflect their modest data requirements. What Can We Get From 1000 Tokens? A Case Study of Multilingual POS Tagging For Resource-Poor Languages Long Duong,12 Trevor Cohn,1 Karin Verspoor,1 Steven Bird,1 and Paul Cook1 1Department of Computing and Information Systems, The University of Melbourne 2National ICT Australia, Victoria Research Laboratory lduong@student.unimelb.edu.au {t.cohn, karin.verspoor, sbird, paulcook}@unimelb.edu.au Abstract In this paper we address the problem of multilingual part-of-speech tagging for resource-poor languages. We use par- allel data to transfer part-of-speech in- formation from resource-rich to resource- poor languages. Additionally, we use a small amount of annotated data to learn to “correct” errors from projected approach such as tagset mismatch between lan- guages, achieving state-of-the-art perfor- mance (91.3%) across 8 languages. Our approach is based on modest data require- ments, and uses minimum divergence clas- sification. For situations where no uni- versal tagset mapping is available, we propose an alternate method, resulting in state-of-the-art 85.6% accuracy on the resource-poor language Malagasy. 1 Introduction Part-of-speech (POS) tagging is a crucial task for natural language processing (NLP) tasks, provid- ing basic information about syntax. Supervised POS tagging has achieved great success, reach- ing as high as 95% accuracy for many languages (Petrov et al., 2012). However, supervised tech- niques need manually annotated data, and this is either lacking or limited in most resource- poor languages. Fully unsupervised POS tagging is not yet useful in practice due to low accu- racy (Christodoulopoulos et al., 2010). In this pa- per, we propose a semi-supervised method to nar- row the gap between supervised and unsupervised approaches. We demonstrate that even a small amount of supervised data leads to substantial im- provement. Our method is motivated by the availability of parallel data. Thanks to the development of mul- tilingual documents from government projects, book translations, multilingual websites, and so forth, parallel data between resource-rich and resource-poor languages is relatively easy to ac- quire. This parallel data provides the bridge that permits us to transfer POS information from a resource-rich to a resource-poor language. Systems that make use of cross-lingual tag projection typically face several issues, includ- ing mismatches between the tagsets used for the languages, artifacts from noisy alignments and cross-lingual syntactic divergence. Our approach compensates for these issues by training on a small amount of annotated data on the target side, demonstrating that only 1k tokens of annotated data is sufficient to improve performance. We first tag the resource-rich language using a supervised POS tagger. We then project POS tags from the resource-rich language to the resource- poor language using parallel word alignments. The projected labels are noisy, and so we use various heuristics to select only “good” training examples. We train the model in two stages. First, we build a maximum entropy classifier T on the (noisy) projected data. Next, we train a supervised classifier P on …

Content truncated to 500 words