Not just dissimilar, but opposite: An algorithm for measuring similarity and oppositeness between words

Many Natural Language Processing applications are underpinned by word comparison algorithms. Often, such comparison algorithms have been designed to determine how related given words are or how similar they are. In some applications, however, it may be important to distinguish between simply not being similar and oppositeness. In this paper we present work in progress on an algorithm that accounts for the similarity or oppositeness of given words. We evaluate this algorithm against `gold-standard' datasets for evaluating similarity and antonymy and discuss some of the issues with these datasets for evaluating algorithms such as ours. We show that our algorithm, using knowledge from WordNet only, is able to achieve precision scores upwards of 84% on the GRE Most Contrasting Word antonymy task, while outperforming state of the art algorithms over the recent SimLex-999 similarity dataset.