Detecting Multiword Expressions by Dependency Parsing

In this poster, we present how different types of MWEs can be identified by dependency parsers in different languages. In our investigations, we focus on English verb-particle constructions (VPCs), Hungarian light verb constructions (LVCs) and German light verb constructions. In our experiments, we exploit the fact that some treebanks contain MWE-aware annotations, i.e. there are MWE-specific morphological or syntactic tags in them. For instance, the French Treebank contains explicit annotations for MWEs (Abeillé et al. 2003) and different version of the Turkish Treebank are also annotated for MWEs (Eryiğit et al. 2011). Here, we make use of the Penn Treebank (Marcus et al., 1993), which contains annotation for VPCs, the TIGER corpus (Brants et al. 2004) and the Szeged Dependency Treebank (Vincze et al. 2013), both of which contain annotation for LVCs. In these treebanks, the special relation of the two components of the MWE is distinctively marked by a certain syntactic label. This entails that if a data-driven syntactic parser is trained on a dataset annotated with extra information for MWEs, it will be able to assign such tags as well, in other words, the syntactic parser itself will be able to identify MWEs in texts. In our experiments, we investigate the performance of such dependency parsers for three languages and two different MWE types.