Dependency Parsing for Identifying Hungarian Light Verb Constructions

International Joint Conference on Natural Language Processing, pages 207–215, Nagoya, Japan, 14-18 October 2013. Dependency Parsing for Identifying Hungarian Light Verb Constructions Veronika Vincze1,2, J´anos Zsibrita2 and Istv´an Nagy T.2 1Hungarian Academy of Sciences, Research Group on Artificial Intelligence vinczev@inf.u-szeged.hu 2Department of Informatics, University of Szeged {zsibrita,nistvan}@inf.u-szeged.hu Abstract Light verb constructions (LVCs) are verb and noun combinations in which the verb has lost its meaning to some degree and the noun is used in one of its original senses. They often share their syntactic pattern with other constructions (e.g. verb- object pairs) thus LVC detection can be viewed as classifying certain syntactic pat- terns as light verb constructions or not. In this paper, we explore a novel way to detect LVCs in texts: we apply a depen- dency parser to carry out the task. We present our experiments on a Hungarian treebank, which has been manually anno- tated for dependency relations and light verb constructions. Our results outper- formed those achieved by state-of-the-art techniques for Hungarian LVC detection, especially due to the high precision and the treatment of long-distance dependencies. 1 Introduction Multiword expressions (MWEs) are lexical items that can be decomposed into single words and dis- play lexical, syntactic, semantic, pragmatic and/or statistical idiosyncrasy (Kim, 2008). Light verb constructions (LVCs) form a subtype of MWEs: they are verb and noun combinations in which the verb has lost its meaning to some degree and the noun is used in one of its original senses (e.g. make a decision or take a walk). In several NLP applica- tions like information retrieval or machine trans- lation it is important to identify LVCs in context since they require special treatment, particularly because of their semantic features. Thus, LVCs should be identified to help these applications. Light verb constructions (e.g. make a mistake) often share their syntactic pattern with literal verb + noun combinations (e.g. make a cake). Thus, specific syntactic constructions – e.g. verb-object pairs – can be separated into two classes: one where the noun behaves as a real object (cake) and one where the noun functions as the light verb object (mistake). Thus, LVC detection can be viewed as classifying certain syntactic patterns as LVCs or not and assigning a specific syntactic label to the argument of the light verb. In this paper, we explore a novel way to LVC detection: we apply a dependency parser to carry out the task. Although the usability of identified multiword expressions has been investigated in the literature (see Section 4), and many MWE detec- tion systems rely on syntactic information, we are not aware of any approach that aimed at apply- ing a dependency parser for the dedicated task of identifying LVCs. Our approach requires a tree- bank annotated for syntactic and LVC informa- tion at the same time. Due to the availability of annotated resources, we focus on light verb con- structions in Hungarian, a morphologically rich language. Thus, we present our experiments on the legal subcorpus of the Szeged Dependency Treebank annotated …

Content truncated to 500 words