Compositionality for perceptual classification

We compare three approaches to compositional semantics for words that are modeled (in part) using perceptual classifiers: ”meanings as sets”, ”meanings as transparent functions”, and ”meanings as opaque functions”. The approaches are evaluated according to whether they fulfill a set of desiderata, including dealing with non-intersective compositionality, working with state of the art classifiers, and accounting for learning of perceptual meanings. We find that none of the current approaches fulfill all our desiderata.