Domain-Dependent Identification of Multiword Expressions

Proceedings of Recent Advances in Natural Language Processing, pages 622–627, Hissar, Bulgaria, 12-14 September 2011. Domain-Dependent Identification of Multiword Expressions Istv´an Nagy T.1, Veronika Vincze2 and G´abor Berend1 1Department of Informatics, University of Szeged {nistvan,berendg}@inf.u-szeged.hu 2Hungarian Academy of Sciences, Research Group on Artificial Intelligence vinczev@inf.u-szeged.hu Abstract The identification of different kinds of multiword expressions require different solutions, on the other hand, there might be domain-related differences in their fre- quency and typology. In this paper, we show how our methods developed for identifying noun compounds and light verb constructions can be adapted to dif- ferent domains and different types of texts. Our results indicate that with little effort, existing solutions for detecting multiword expressions can be successfully applied to other domains as well. 1 Introduction Multiword expressions (MWEs) are lexical units that consist of more than one orthographical word, i.e. a lexical unit that contains spaces (Sag et al., 2002; Calzolari et al., 2002). There are several methods developed for identifying several types of MWEs, however, different kinds of multiword ex- pressions require different solutions. Furthermore, there might be domain-related differences in the frequency of a specific MWE type. In this paper, we show how our methods developed for identify- ing noun compounds and light verb constructions can be adapted to different domains and different types of texts, namely, Wikipedia articles and texts from various topics. Our results suggest that with simple modifications, competitive results can be achieved on the target domains. 2 Related work There are several solutions developed for identify- ing different types of MWEs in different domains. Bonin et al. (2010) use contrastive filtering in or- der to identify multiword terminology in scien- tific, Wikipedia and legal texts: term candidates are ranked according to their belonging to the gen- eral language or the sublanguage of the domain. The tool mwetoolkit (Ramisch et al., 2010a) is designed to identify several types of MWEs in different domains, which is illustrated by identi- fying English compound nouns in the Genia and Europarl corpora and in general texts (Ramisch et al., 2010b; Ramisch et al., 2010c). Statistical models are used for the identification of several types of multiword expressions in sev- eral languages (e.g. Bouma (2010),Villavicencio et al. (2007)). However, they require (costly) an- notated resources on the one hand and they are not able to identify rare MWEs in corpora on the other hand – as Piao et al. (2003) emphasize, about 68% of multiword expressions occur only once or twice in their corpus. Some hybrid systems make use of both statis- tical and linguistic information as well, that is, rules based on syntactic or semantic regularities are also incorporated into the system (Bannard, 2007; Cook et al., 2007; Al-Haj and Wintner, 2010). This results in better coverage of multi- word expressions. On the other hand, these meth- ods are highly specific because of the amount of linguistic rules encoded, thus, it requires much ef- fort to adapt them to different languages or even to different types of multiword expressions. …

Content truncated to 500 words