Deep Lexical Acquisition of Type Properties in Low-resource Languages: A Case Study in Wambaya

Copyright 2012 by Jeremy Nicholson, Rachel Nordlinger, and Timothy Baldwin 26th Pacific Asia Conference on Language,Information and Computation pages 481–490 Deep Lexical Acquisition of Type Properties in Low-resource Languages: A Case Study in Wambaya Jeremy Nicholson,†‡ Rachel Nordlinger‡ and Timothy Baldwin†‡ † NICTA Victoria Research Laboratories ‡ The University of Melbourne, VIC 3010, Australia {nj,racheln,tbaldwin}@unimelb.edu.au Abstract We present a case study on applying com- mon methods for the prediction of lexical properties to a low-resource language, namely Wambaya. Leveraging a small corpus leads to a typical high-precision, low-recall system; using the Web as a corpus has no utility for this language, but a machine learning approach seems to utilise the available resources most effectively. This motivates a semi-supervised approach to lexicon extension. 1 Introduction Deep lexical acquisition (DLA) is the pro- cess of (semi-)automatically creating or extend- ing linguistically-rich lexical resources (Baldwin, 2005a; Baldwin, 2005b; Baldwin, 2007). Conven- tionally, DLA has been applied to high-resourced languages such as English, German or Japanese to broaden the coverage of a medium-coverage re- source, or enrich existing linguistic annotation in resources. However, it also has tremendous poten- tial in accelerating the documentation of low-density languages, a fact that is often discussed but very rarely delivered on in the literature. This paper at- tempts to deliver on this promise, and asks the ques- tion: how well do standard approaches to DLA per- form over low-density languages? For example, one of the standard approaches to DLA is to extract n-gram counts for patterns involving a target lex- eme from the web, and use these as the basis for predicting the lexical class membership of the lex- eme. While there is little expectation that we will find significant amounts of text for low-density lan- guages on the web, we nevertheless run the experi- ment to test the general applicability of this style of approach. In this work, we take Wambaya as a real-world chronically low-density language, and examine the task of predicting the grammatical gender of nomi- nal lexical items. Wambaya is a nearly extinct lan- guage (Gordon, 2005) from the Mirndi group of Australian languages. Like many languages from the Australian family, its complicated syntax and rich morphology makes natural language process- ing of Wambaya difficult. Unlike many of its neigh- bours, however, it has been well documented in a descriptive grammar (Nordlinger, 1998) and a Head- driven Phrase Structure Grammar (Bender, 2008). While resources for Wambaya are of little intrinsic value as it is doubtful that new text will be gener- ated in the language, developing these resources is still instructive for parallel development in compara- ble languages (Warlpiri, for example, has a notable speech community (Gordon, 2005)). Additionally, it provides an invaluable test bed for DLA research, to test the potential of methods over similarly low- density languages, and truly test the bounds of DLA for the purposes of language preservation. Lexicon extension for Wambaya is a task compa- rable to the state of many resources: the available lexicon is small, of only …

Content truncated to 500 words