Exploring Vector Space Models to Predict the Compositionality of German Noun-Noun Compounds

This paper explores two hypotheses regarding vector space models that predict the compositionality of German noun-noun compounds: (1) Against our intuition, we demonstrate that window-based rather than syntax-based distributional features perform better predictions, and that not adjectives or verbs but nouns represent the most salient part-of-speech. Our overall best result is state-of-the-art, reaching Spearman’s ρ = 0.65 with a wordspace model of nominal features from a 20word window of a 1.5 billion word web corpus. (2) While there are no significant differences in predicting compound–modifier vs. compound–head ratings on compositionality, we show that the modifier (rather than the head) properties predominantly influence the degree of compositionality of the compound.