Modelling Polysemy in Adjective Classes by Multi-Label Classification

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 171–180, Prague, June 2007. c⃝2007 Association for Computational Linguistics Modelling Polysemy in Adjective Classes by Multi-Label Classification Gemma Boleda GLiCom Universitat Pompeu Fabra 08003 Barcelona gemma.boleda@upf.edu Sabine Schulte im Walde IMS University of Stuttgart 70174 Stuttgart schulte@ims. uni-stuttgart.de Toni Badia GLiCom Universitat Pompeu Fabra 08003 Barcelona toni.badia@upf.edu Abstract This paper assesses the role of multi-label classification in modelling polysemy for lan- guage acquisition tasks. We focus on the ac- quisition of semantic classes for Catalan ad- jectives, and show that polysemy acquisition naturally suits architectures used for multi- label classification. Furthermore, we ex- plore the performance of information drawn from different levels of linguistic descrip- tion, using feature sets based on morphol- ogy, syntax, semantics, and n-gram distribu- tion. Finally, we demonstrate that ensemble classifiers are a powerful and adequate way to combine different types of linguistic ev- idence: a simple, majority voting ensemble classifier improves the accuracy from 62.5% (best single classifier) to 84%. 1 Introduction This paper reports on a series of experiments to ex- plore the automatic acquisition of semantic classes for Catalan adjectives. The most important chal- lenge of the classification task is to model the assign- ment of polysemous lexical instances to multiple se- mantic classes, combining a) a state-of-the-art Ma- chine Learning architecture for Multi-label Classi- fication (Schapire and Singer, 2000; Ghamrawi and McCallum, 2005) and an Ensemble Classifier (Di- etterich, 2002) with b) the definition of features at various levels of linguistic description. A proper treatment of polysemy is essential in the area of lexical acquisition, since polysemy repre- sents a pervasive phenomenon in natural language. However, previous approaches to the automatic ac- quisition of semantic classes have mostly disre- garded the problem (cf. Merlo and Stevenson, 2001 and Stevenson and Joanis, 2003 for English seman- tic verb classes, or Schulte im Walde, 2006 for Ger- man semantic verb classes). There are a few excep- tions to this tradition, such as Pereira et al. (1993), Rooth et al. (1999), Korhonen et al. (2003), who used soft clustering methods for multiple assign- ment to verb semantic classes. Our work addresses the lack of methodology in modelling a polysemous classification. We imple- ment a multi-label classification architecture to han- dle polysemy. This paper concentrates on the clas- sification of Catalan adjectives, but the general na- ture of the architecture should allow related tasks to profit from our insights. As target classification for the experiments, a set of 210 Catalan adjectives was manually classified by experts into three simple and three polysemous se- mantic classes. We deliberately decided in favour of a small-scale, broad classification. So far, there is little work on the semantic classification of adjec- tives, as opposed to verbal semantic classification. The semantic classification we propose is a first step in characterising adjectival meaning, and can be re- fined and extended in subsequent work. The experiments also provide a thorough compar- ison of feature …

Content truncated to 500 words