Interpretable features for meaning in context

Recently, interpretable vector spaces have been introduced, vector spaces in which individual dimensions are meaningful. For polysemous words, interpretable vectors mix features from different senses. In this paper, we propose the task of computing contextspecific interpretable representations for individual senses of a word. Focusing on computing features that describe circumstances of an event, we find that the task is feasible: The best vector contextualization model, component-wise multiplication, achieves positive ratings in 83% of all cases, compared to 58% for a baseline model.