igk4-Hutchinson.html

For word classes such as nouns, humans agree fairly well on which lexical items are semantically similar. Furthermore, if two nouns tend to appear in similar contexts, they are likely to be semantically similar as well (Rubenstein and Goodenough, 1965; Miller and Charles, 1991). This talk addresses the question of whether similar results hold for discourse connectives, e.g. 'whereas', 'because', 'only if'. These connectives introduce relations between abstract objects such as events, propositions and speech acts, and judging their similarity is arguably more difficult than it is for nouns. In our first experiment we determine whether or not subjects agree on similarity of discourse connectives, and relate similarity judgements to the substitutability of the connectives. In our second experiment, we model the human similarity judgements using measures of distributional similarity.