Unsupervised Word Class Induction for Under-resourced Languages: A Case Study on Indonesian

International Joint Conference on Natural Language Processing, pages 685–691, Nagoya, Japan, 14-18 October 2013. Unsupervised Word Class Induction for Under-resourced Languages: A Case Study on Indonesian Meladel Mistica The Australian National University meladel.mistica@gmail.com Jey Han Lau and Timothy Baldwin The University of Melbourne jeyhan.lau@gmail.com tb@ldwin.net Abstract In this study we investigate how we can learn both: (a) syntactic classes that cap- ture the range of predicate argument struc- tures (PASs) of a word and the syntactic alternations it participates in, but ignore large semantic differences in the compo- nent words; and (b) syntactico-semantic classes that capture PAS and alternation properties, but are also semantically co- herent (a la Levin classes). We focus on Indonesian as our case study, a language that is spoken by more than 165 million speakers, but is nonetheless rela- tively under-resourced in terms of NLP. In particular, we focus on the syntactic varia- tion that arises with the affixing of the In- donesian suffix -kan, which varies accord- ing to the kind of stem it attaches to. 1 Introduction This research was motivated by the desire to semi-automatically develop a lexicon for a wide- coverage, precision grammar of Indonesian. Al- though these linguistically-motivated grammars are invaluable resources to the NLP community, the biggest drawback is the time required for the manual creation and curation of the lexicon. Our work aims to expedite this process by automati- cally assigning syntactic information to stems that make up the verbal elements, on the basis of pre- dicting syntactico-semantic clusters based on dis- tributional similarity. However, one minor point becomes one ma- jor obstacle in this task: Indonesian is a rela- tively under-resourced language in terms of NLP. Therefore, many of the techniques that have been deemed successful in the inferring of syntactic in- formation or inducing syntactico-semantic classes are not available to us. Even studies that are con- sidered lightweight minimally employ a part-of- speech (POS) tagger and chunker (Joanis et al., 2008), with many studies benefiting from the rich- ness of the features that a syntactic parser pro- vides (Schulte im Walde, 2006). In the case of Indonesian, there exist POS taggers (Pisceldo et al., 2009; Wicaksono and Purwarianti, 2010)1 but no chunker or syntactic parser, and the reliance on such pre-processing tools is unrealistic. We adhere to the notion that semantic similar- ity begets syntactic similarity as per Levin (1989), and so employ a distributional similarity method to learn our syntactic classes, based on a non- parametric Bayesian model. We experiment with learning both: (a) syntactic classes that capture the range of predicate argument structures (PASs) of a word and the syntactic alternations it partic- ipates in, but ignore large semantic differences in the component words; and (b) syntactico-semantic classes that capture PAS and alternation proper- ties, but are also semantically coherent (a la Levin classes). Here, we focus on the syntactic variation that arises with the affixing of the Indonesian suffix -kan. The specific morpho-syntactic behaviour of the kan-affixed verb is very much determined by the …

Content truncated to 500 words