Real-World Semi-Supervised Learning of POS-Taggers for Low-Resource Languages

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 583–592, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Real-World Semi-Supervised Learning of POS-Taggers for Low-Resource Languages Dan Garrette1 Jason Mielens2 1Department of Computer Science 2Department of Linguistics The University of Texas at Austin The University of Texas at Austin dhg@cs.utexas.edu {jmielens,jbaldrid}@utexas.edu Jason Baldridge2 Abstract Developing natural language processing tools for low-resource languages often re- quires creating resources from scratch. While a variety of semi-supervised meth- ods exist for training from incomplete data, there are open questions regarding what types of training data should be used and how much is necessary. We dis- cuss a series of experiments designed to shed light on such questions in the con- text of part-of-speech tagging. We obtain timed annotations from linguists for the low-resource languages Kinyarwanda and Malagasy (as well as English) and eval- uate how the amounts of various kinds of data affect performance of a trained POS-tagger. Our results show that an- notation of word types is the most im- portant, provided a sufficiently capable semi-supervised learning infrastructure is in place to project type information onto a raw corpus. We also show that finite- state morphological analyzers are effective sources of type information when few la- beled examples are available. 1 Introduction Low-resource languages present a particularly dif- ficult challenge for natural language processing tasks. For example, supervised learning meth- ods can provide high accuracy for part-of-speech (POS) tagging (Manning, 2011), but they per- form poorly when little supervision is avail- able. Good results in weakly-supervised tagging have been obtained by training sequence models such as hidden Markov models (HMM) using the Expectation-Maximization algorithm (EM), how- ever most work in this area has still relied on rel- atively large amounts of data, both annotated and unannotated, as well as an assumption that the an- notations are very clean (Kupiec, 1992; Merialdo, 1994). The ability to learn taggers using very little data is enticing: only a tiny fraction of the world’s lan- guages have enough data for standard supervised models to work well. The collection or develop- ment of resources is a time-consuming and expen- sive process, creating a significant barrier for an under-studied language where there are few ex- perts and little funding. It is thus important to develop approaches that achieve good accuracy based on the amount of data that can be reasonably obtained, for example, in just a few hours by a lin- guist doing fieldwork on a non-native language. Previous work explored learning taggers from weak information, but the type, amount, quality, and sources of data raise questions about the appli- cability of those results to real-world low-resource scenarios (Toutanova and Johnson, 2008; Ravi and Knight, 2009; Hasan and Ng, 2009; Garrette and Baldridge, 2012). Most research simulated weak supervision with tag dictionaries extracted from existing large, expertly-annotated corpora. These resources have been developed over long periods of time by trained annotators who collaborate to produce high-quality analyses. They are also bi- ased …

Content truncated to 500 words