This paper deals with the task of finding generally applicable substitutions for a given input term. We show that the output of a distributional similarity system baseline can be filtered to obtain terms that are not simply similar but frequently substitutable. Our filter relies on the fact that when two terms are in a common entailment relation, it should be possible to substitute one for the other in their most frequent surface contexts. Using the Google 5-gram corpus to find such characteristic contexts, we show that for the given task, our filter improves the precision of a distributional similarity.
