The seminal 1973 Cup work of the linguist William Labov demonstrated that denotation - or the association between a set of linguistic labels and a set of possible referents - is subject to the same indeterminacy as other areas of human categorisation. Here, even small variations in features
will contribute to a gestalt judgement of greater or lesser prototypicality for a lexical category, with gradual differences giving way at some point to categorical distinctions. Such issues may be amenable to computational investigation as long as the study is grounded through prototypical exemplars of each category, as judged by human subjects under appropriate conditions.

In this talk, I will consider the problem of categorisation for denotata (here a range of facial images known to be associated with terms for human emotion) through a series of machine learning
studies of the famous Ekman and Friesen ``Pictures of Facial Affect'' images: the domain Gaussian receptive field neural network; individual support vector machine models; and recent work
on combination of multiple SVMs through voting schemes.

In each case, the model is trained upon images derived from the Ekman and Friesen  database, and is able subsequently to generalise successfully to images of unseen subjects. By using digital morphing techniques to produce intermediate frames between the existing stills, we predict a complex range of transitions between denotata, and that such a set may have only a limited role in the acquisition of more complex emotional terms.

(*) By the way: computers can also find Sadness, Surprise, Fear, Disgust, and Anger...