Supervised Approaches for Detection of Non-Compositional Nominal Compounds

Сompositionality of noun compounds indicates to what extent the meaning of a phrase can be derived from the meaning of its parts and their grammatical relations. If semantics of a compound differs from its own components, the phrase is called non-compositional. Some of the examples of non-compositional collocations are "hot dog" and "rat race". This can be applied in various areas, notably, in mashine translation. Detection of compositionality is a task that has been frequently addressed with Distributional Semantic Models (DSMs) and mathematical operations with corresponding word vectors. Unlike traditional approaches that use an "unsupervised setting", we treat this task as a classification problem and introduce various supervised learning algorithms to gain substantial improvement across gold standard dataset over state-of-the-art models. We also proceed to introduce Russian language dataset for compound compositionality task with ADJ+NOUN and NOUN+NOUN speech patterns, the first one among Slavic languages to our knowledge, and perform same classification methods to introduce baseline results comparable with english corpora.