ABSTRACT
This chapter describes an experimental computer program that deduces the meaning of novel verbs from the context of story descriptions. The key idea is a variation on Winston's program that learned the structural descriptions of block world scenes. The word-learning component uses the larger system's determination of the syntactic category of a new word and its predicate-argument structure. A learning theory can be built only if one has a good representation of what it is that is learned, and some idea of the constraints on that target state. The chapter suggests a concrete computational model for the acquisition of word protoypes from positive examples. It reviews two parts of the learning systems behavior that are most open for revision. The first weak point is our choice of representational language, and the second our predefined set of AKO relationships. The second weak point of the learning model is that it depends on a preexisting set of AKO descriptors.
