A Word-Embedding-based Sense Index for Regular Polysemy Representation

We present a method for the detection and representation of polysemous nouns, a phenomenon that has received little attention in NLP. The method is based on the exploitation of the semantic information preserved in Word Embeddings. We first prove that polysemous nouns instantiating a particular sense alternation form a separate class when clustering nouns in a lexicon. Such a class, however, does not include those polysemes in which a sense is strongly predominant. We address this problem and present a sense index that, for a given pair of lexical classes, defines the degree of membership of a noun to each class: polysemy is hence implicitly represented as an intermediate value on the continuum between two classes. We finally show that by exploiting the information provided by the sense index it is possible to accurately detect polysemous nouns in the dataset.