Using a Rich Feature Set for the Identification of German MWEs

Due to the formal variability and the irregular behaviour of MWEs on different levels of linguistic description, they are a potential source of errors for many NLP applications, e.g. Machine Translation. While most of the known approaches to MWE identification focus on one dimension of irregular behaviour, we present an approach that combines morpho-syntactic features (extracted from dependency parsed text) with semantic opacity features (approximated using word alignments). We trained supervised classifiers with different feature sub-sets and show that the combination of morphosyntactic and semantic opacity features yields best overall results.