On the Compositionality and Semantic Interpretation of English Noun Compounds

In this paper we present a study covering the creation of compositional distributional representations for English noun compounds (e.g. computer science) using two compositional models proposed in the literature. The compositional representations are first evaluated based on their similarity to the corresponding corpus-learned representations and then on the task of automatic classification of semantic relations for English noun compounds. Our experiments show that compositional models are able to build meaningful representations for more than half of the test set compounds. However, using pre-trained compositional models does not lead to the expected performance gains for the semantic relation classification task. Models using compositional representations have a similar performance as a basic classification model, despite the advantage of being pretrained on a large set of compounds.