Opportunities and challenges in semantic similarity

Semantic similarity has been increasingly adopted in the recent past as a viable, scalable alternative to the full-understanding approach to natural language understanding. We present here an overview of opportunities and challenges to semantic similarity with an emphasis on methods, data, and tools. A series of methods we developed over the past decade will be summarized. These methods and others have been integrated in a semantic similarity toolkit called SEMILAR (www.semanticsimilarity.org) which has been widely adopted by thousands of users since its launch in summer of 2013 at the Annual Meeting of the Association for Computational Linguistics. Furthermore, we will illustrate some drawbacks of current data sets that hamper a fair comparison among existing methods. Several suggestions will be made to improve the building of future data sets for assessing approaches to semantic similarity