Noun compound semantics: Linguistic and general -purpose reasoning.
Mark Arehart · Deep Blue (University of Michigan) · 2003
Noun compounds resist standard compositional semantic analysis because the modification relation is implicit and because nouns have been thought to lack argument structure. Early analyses provided argument positions for nouns using phrase structure rules that licensed phonologically null predicates. Recent work proposes detailed nominal argument structure and enriched semantic composition. Despite the shift in emphasis from syntax to the lexicon, the prevailing approaches still encode meanings grammatically. This dissertation offers three main contributions to the semantic analysis and processing of compounds. First, I problematize the traditional distinctions made in analyses of compounds, beginning with the strict separation of compounds that have deverbal heads from those with non-deverbal heads. I argue that these classes participate in the same sorts of modification relations. I also present a novel classification based on the logical structure of denotations, and I offer independent motivation for the classification through an analysis of word-order constraints. The second main contribution is to recast interpretation as a process of general-purpose reasoning rather than linguistic rule application. The third contribution is a model that uses measures of concept and word probabilities and inferencing procedures utilizing a general ontology. The denotations generated by the system are used in several tasks: disambiguation among semantic classes, as input for generating informative paraphrases, and as queries to select potential referents in a 3D graphical environment. The disambiguation task tests the system's ability to make certain semantic distinctions regarding the type of modification and the logical form of the denotation. The paraphrasing task allows naive judges to rate paraphrases based on denotations that are generated using different degrees of intelligence available in the system. The results generally demonstrate above-baseline performance, suggesting that the approach represents a fruitful combination of symbolic and statistical methods for generating and ranking denotations.