Detecting Aspectual Relations Quantitatively

The explosion in the creation of text corpora in recent years suggests that the opportunity may be ripe to examine quantitative techniques for their value in semantics. The present paper aims to explore one quantitative technique with an eye toward its potential value in illuminating questions of semantic theory. It is exploratory in nature, and it will not offer definite conclusions or specific advice. To exploit large text corpora—of a size of the order of magnitude of 10 9 words, we need to employ automatic procedures of analysis. It is unthinkable to work through such volumes of material except by using computer programs. This leads to a difficulty when one's analytical ambitions are semantic, since the semantics of texts is not immediately accessible to automatic procedures. Perhaps in some glorious future there will be data annotated for semantics to an extent that makes the direct application of quantitative analyses straightforward. But at present we do not have such resources, only fairly large corpora of texts. This means in turn that we need to operationalize our semantic concepts in a way that is amenable to automatic processing. Naturally this has an impact on the sorts of phenomena that can be studied. In the present study we concentrate on aspectual categories and on the affinities different sorts of adverbials have with some aspects as opposed to others. We operationalize the inherent aspect of verb phrases by noting the verb heading the phrase, and the class of the durative adverbial by noting the preposition heading the durative adverbial. Both of these steps are subject to some error, but fortunately, quantitative techniques also tolerate error in characterizations as long as the statistical 3 4 / Detecting Aspectual Relations Quantitatively strength of the association between the category and its operationaliza-tion is sufficient. We discuss computational semantics in this section as well as the statistical background needed for detecting aspectual affinities in large corpora. 1.2.1 Computational Semantics Work in computational linguistics has frequently engaged work in semantic theory essentially by implementing and applying ideas from the latter, often with special interest for issues concerning disambigua-tion (Nerbonne, 1996). Since the quantitative turn in computational linguistics lexical semantics have been the focus of attention in computational work. Even though it is difficult to collect the statistical information that is useful to lexical semantics, many quantitative techniques have been proposed with an eye toward detecting semantic properties. For example, …