Seeing Arguments through Transparent Structures

This paper describes a research effort that exploits information available in the FrameNet database and seeks to find, for argumentstructure- bearing verbs, nouns, and adjectives, the lexical heads of the phrases that satisfy the core semantic roles of those predicates, and to create from the database of annotated sentences collections of structured clusters of words, called kernel dependency graphs. These KDGs when thus extracted from a collection of annotated sentences can be studied as candidates for the status of special collocations, but the same kinds of clusters, when discovered in raw text, can serve, in NLP applications, as indicators of the salient topics or issues in the passage from which they have been extracted. Unfortunately, there are sometimes discrepancies between syntactic and semantic heads , and for our purposes it is the semantic head that is significant; it is thus necessary to identify grammatical structures - and the words which mark them - that can intervene, structurally, between a predicate and its arguments. When these transparency structures are the familiar control structures seen in various kinds of embedding predicates, we should be able to rely on ordinary parsers to identify them; but the concern in this paper is with two additional phenomena, the support verbs that separate arguments from predications that are expressed as nouns (typically deverbal nouns), and transparent nouns that syntactically take the semantically relevant nouns as their complements
Published in 2002