Meanings \ from Examples

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 (Winston, 1975) program that learned the structural descriptions of block world scenes. Instead of learning descriptions of toy block assemblies like ARCH and TOWER, the word-learning program acquires frame-based descriptions of English verbs like MURDER or DONATE. The program works by assuming that similar verbs will play similar causal roles in common story plots. For example. ASSASSINATE is like MURDER because both MURDER and ASSASSINATE cause similar things to happen and are caused by similar patterns and events. Intuitively, we learn about a new verb like ASSASSINATE as a kind of family resemblance variation on a core verb like MURDER. The program works in a similar way. Syntactic constraints derived from the parsing of story plots are used to drive an analogical matching procedure. Analogical matching gives a way to compare descriptions of known words to unknown words. The "meaning" of a new verb is learned by matching part of the causal network description of a story precis