Idiomatic expression usage recognition by neural networks

Many of the idiomatic expressions can be used both in literal and non-literal ways. The recognition of such cases is an important problem in many natural language processing applications, namely, in machine translation. We propose automatic idiom usage recognition method based on the analysis of local contexts of such expressions. We apply recurrent neural networks to solve this problem. Two types of neural networks are investigated - simple and bidirectional recurrent networks. We compare two forms of representation of context words - the canonical form (by lemmas) and by source word forms. We describe construction and parameters of the distributive model which stores the vector representations of single words and target idiomatic expressions. Due to the great diversity of approaches to solving the idiom usage recognition problem, we provide an extended survey of basic efforts in this domain.