Die Morphologie (f): Targeted Lexical Acquisition for Languages other than English

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 67–74. Die Morphologie (f) : Targeted Lexical Acquisition for Languages other than English Jeremy Nicholson†, Timothy Baldwin†, and Phil Blunsom‡ †‡ Department of Computer Science and Software Engineering University of Melbourne, VIC 3010, Australia and † NICTA Victoria Research Laboratories University of Melbourne, VIC 3010, Australia {jeremymn,tim,pcbl}@csse.unimelb.edu.au Abstract We examine standard deep lexical acqui- sition features in automatically predict- ing the gender of noun types and to- kens by bootstrapping from a small an- notated corpus. Using a knowledge-poor approach to simulate prediction in unseen languages, we observe results comparable to morphological analysers trained specif- ically on our target languages of German and French. These results describe fur- ther scope in analysing other properties in languages displaying a more challenging morphosyntax, in order to create language resources in a language-independent man- ner. 1 Introduction As a result of incremental annotation efforts and advances in algorithm design and statistical mod- elling, deep language resources (DLRs, i.e. lan- guage resources with a high level of linguistic sophistication) are increasingly being applied to mainstream NLP applications. Examples include analysis of semantic similarity through ontologies such as WordNet (Fellbaum, 1998) or VerbOcean (Chklovski and Pantel, 2004), parsing with preci- sion grammars such as the DELPH-IN (Oepen et al., 2002) or PARGRAM (Butt et al., 2002) gram- mars, and modelling with richly annotated tree- banks such as PropBank (Palmer et al., 2005) or CCGbank (Hockenmaier and Steedman, 2002). Unfortunately, the increasing complexity of these language resources and the desire for broad coverage has meant that their traditionally man- ual mode of creation, development and mainte- nance has become infeasibly labour-intensive. As a consequence, deep lexical acquisition (DLA) has been proposed as a means of automatically learn- ing deep linguistic representations to expand the coverage of DLRs (Baldwin, 2005). DLA research can be divided into two main categories: targeted DLA, in which lexemes are classified according to a given lexical property (e.g. noun countability, or subcategorisation properties); and generalised DLA, in which lexemes are classified according to the full range of lexical properties captured in a given DLR (e.g. the full range of lexical relations in a lexical ontology, or the system of lexical types in an HPSG). As we attest in Section 2, most work in deep lexical acquisition has focussed on the English language. This can be explained in part by the ready availability of targeted language resources like the ones mentioned above, as well as sec- ondary resources — such as corpora, part-of- speech taggers, chunkers, and so on — with which to aid prediction of the lexical property in ques- tion. One obvious question, therefore, is whether the techniques used to perform lexical acquisition in English generalise readily to other languages, where subtle but important differences in mor- phosyntax might obfuscate the surface cues used for prediction. In this work, we will examine the targeted pre- diction of the gender of noun types and tokens in context, for both German and French. …

Content truncated to 500 words