This paper aims to re-think the role of the word similarity task in distributional semantics research. We argue while it is a valuable tool, it should be used with care because it provides only an approximate measure of the quality of a distributional model. Word similarity evaluations assume there exists a single notion of similarity that is independent of a particular application. Further, the small size and low inter-annotator agreement of existing data sets makes it challenging to find significant differences between models.

No abstract