Learning English Light Verb Constructions: Contextual or Statistical

In this paper, we investigate a supervised machine learning framework for automatically learning of English Light Verb Constructions (LVCs). Our system achieves an 86.3% accuracy with a baseline (chance) performance of 52.2% when trained with groups of either contextual or statistical features. In addition, we present an in-depth analysis of these contextual and statistical features and show that the system trained by these two types of cosmetically different features reaches similar performance empirically. However, in the situation where the surface structures of candidate LVCs are identical, the system trained with contextual features which contain information on surrounding words performs 16.7% better. In this study, we also construct a balanced benchmark dataset with 2,162 sentences from BNC for English LVCs. And this data set is publicly available and is also a useful computational resource for research on MWEs in general.