Indonesian Dependency Treebank: Annotation and Parsing

Copyright 2012 by Nathan Green, Septina Dian Larasati, and Zdenek Zabokrtsky 26th Pacific Asia Conference on Language,Information and Computation pages 137–145 Indonesian Dependency Treebank: Annotation and Parsing Nathan Green1, Septina Dian Larasati1,2 and Zdenˇek ˇZabokrtsk´y1 1Charles University in Prague 2SIA Tilde Institute of Formal and Applied Linguistics Vienibas gatve 75a Faculty of Mathematics and Physics LV-1004 Prague, Czech Republic Riga, Latvia {green,larasati,zabokrtsky}@ufal.mff.cuni.cz Abstract We introduce and describe ongoing work in our Indonesian dependency treebank. We de- scribed characteristics of the source data as well as describe our annotation guidelines for creating the dependency structures. Reported within are the results from the start of the In- donesian dependency treebank. We also show ensemble dependency pars- ing and self training approaches applicable to under-resourced languages using our man- ually annotated dependency structures. We show that for an under-resourced language, the use of tuning data for a meta classi- fier is more effective than using it as ad- ditional training data for individual parsers. This meta-classifier creates an ensemble de- pendency parser and increases the dependency accuracy by 4.92% on average and 1.99% over the best individual models on average. As the data sizes grow for the the under-resourced language a meta classifier can easily adapt. To the best of our knowledge this is the first full implementation of a dependency parser for In- donesian. Using self-training in combination with our Ensemble SVM Parser we show adi- tional improvement. Using this parsing model we plan on expanding the size of the corpus by using a semi-supervised approach by applying the parser and correcting the errors, reducing the amount of annotation time needed. 1 Introduction Treebanks have been a major source for the advance- ment of many tools in the NLP pipeline from sen- tence alignment to dependency parsers to an end product, which is often machine translation. While useful for machine learning as well and linguistic analysis, these treebanks typically only exist for a handful of resource-rich languages. Treebanks tend to come in two linguistic forms, dependency based and constituency based each with their own pros and cons. Dependency treebanks have been made pop- ular by treebanks such as the Prague dependency treebank (Hajic, 1998) and constituency treebanks by the Penn treebank (Marcus et al., 1993). While some linguistic phenomena are better represented in one form instead of another, the two forms are gen- erally able to be transformed into one another. While many of the world’s 6,000+ languages could be considered under-resourced due to a lim- ited number of native speakers and low overall popu- lation in their countries, Indonesia is the fourth most populous country in the world with over 23 million native and 215 million non-native Bahasa Indonesia speakers. The development of language resources, treebanks in particular, for Bahasa Indonesia will have an immediate effect for Indonesian NLP. Further development of our Indonesian depen- dency treebank can affect part of speech taggers, named entity recognizers, and machine translation systems. All of these systems have technical bene- fits to the 238 …

Content truncated to 500 words