Adaptability of Lexical Acquisition for Large-scale Grammars

Proceedings of Recent Advances in Natural Language Processing, pages 355–362, Hissar, Bulgaria, 12-14 September 2011. Adaptability of Lexical Acquisition for Large-scale Grammars Kostadin Cholakov†, Gertjan van Noord†, Valia Kordoni‡, Yi Zhang‡ † University of Groningen, The Netherlands ‡ Saarland University and DFKI GmbH, Germany {k.cholakov,g.j.m.van.noord}@rug.nl {kordoni,yzhang}@coli.uni-saarland.de Abstract In this paper, we demonstrate the portabil- ity of the lexical acquisition (LA) method proposed in Cholakov and van Noord (2010a). Here, LA refers to the acqui- sition of linguistic descriptions for words which are not listed in the lexicon of a given computational grammar, i.e., words which are unknown to this grammar. The method we discuss was originally devel- oped for the Dutch Alpino system, and the paper shows that the method also applies to the GG (Crysmann, 2003), a compu- tational HPSG grammar of German. The LA method obtains very similar results for German (84% F-measure on learning un- known words). Extending the GG with the lexical entries proposed by the LA method causes an important improvement in pars- ing accuracy for a test set of sentences con- taining unknown words. Furthermore, in a smaller experiment, we show that the linguistic knowledge the LA method pro- vides can also be used for sentence gener- ation. 1 Introduction Computational grammars of natural language lie at the heart of various wide-coverage symbolic parsing systems. At present, such systems have been integrated into real-world NLP applications, such as IE, QA, grammar checking, MT and intel- ligent IR. This integration, though, has reminded us of some of the problems which the aforemen- tioned grammars encounter when applied to nat- urally occurring text, in particular lack of lexi- cal coverage. Since such grammars usually rely on hand-crafted lexicons containing elaborate lin- guistic descriptions, words not listed in the lexi- con, i.e. words unknown to the grammar, pose a major issue in the employment of the grammars for real-life applications. In this context, lexical acquisition refers to the acquisition of correct lex- ical descriptions for unknown words. Various LA techniques for computational gram- mars have been proposed in the past. Cussens and Pulman (2000) used a symbolic approach em- ploying inductive logic programming, while Er- bach (1990), Barg and Walther (1998) and Fou- vry (2003) followed a unification-based approach. Other approaches have treated LA as a classifi- cation task where the unknown word is mapped to a finite set of labels. Baldwin (2005) has ex- tracted features from various linguistic resources (POS taggers, chunkers, etc.) and used a set of bi- nary classifiers to learn lexical entries for a large- scale grammar of English (ERG; (Copestake and Flickinger, 2000)). Zhang and Kordoni (2006) and Cholakov et al. (2008), on the other hand, have trained a maximum entropy (ME) classifier with features extracted from the grammar in order to ac- quire new lexical entries for the ERG and the GG (Crysmann, 2003), respectively. Extending this line of research, Cholakov and van Noord (2010a) have proposed a technique for learning unknown words for the Dutch Alpino grammar (van Noord, …

Content truncated to 500 words