Sentiment dictionary refinement using word embeddings

Abstract

Previous works on Polish sentiment dictionaries revealed the superiority of machine learning on vectors created from word contexts (concordances or word co-occurrence distributions), especially compared to the SO-PMI method (semantic orientation of pointwise mutual information). This paper demonstrates that this state-of-the-art method could be improved upon when extending the vectors by word embeddings, obtained from skip-gram language models. Specifically, it proposes a new method of computing word sentiment polarity using feature sets composed of vectors created from word embeddings and word co-occurrence distributions. The new technique is evaluated in a number of experimental settings.