Citation Context

... [9] that may need weeks to train. Distributed word representations, also known as word embeddings, represent a word as a vector in Rn. The more the vectors are closer the more the two corresponding words are deemed to share some degree of syntactical or semantical similarity. Word embeddings are typically obtained as the by-product of the training of a neural language model [6,10] that learns to predict the most probable word given a set of context words. Word embeddings have empirically proved to be helpful in the solution on several tasks that span from NLP [17,21] to information retrieval [3,8]. In this work we compare the performance on the Italian language of two widely adopted “shallow” word representation models, word2vec and GloVe [18]. We train them on two datasets that have different stylistic properties, with the aim of checking if such differences have an impact on the produced embeddings. We have produced an Italian version of the Google word analogy test in order qualitatively compare the performance of the different models, we make it publicly available. We also make publicly available the word vectors obtained by the above-mentioned models on the two datasets1. 2 Relate...