Abstract

Written communication is rarely a sequence of simple assertions. More often, in addition to simple assertions, authors express subjectivity, such as beliefs, speculations, opinions, intentions, and desires. Furthermore, they link statements of various kinds to form a coherent discourse that reflects their pragmatic intent. In computational semantics, extraction of simple assertions (propositional meaning) has attracted the greatest attention, while research that focuses on extra-propositional aspects of meaning has remained sparse overall and has been largely limited to narrowly defined categories, such as hedging or sentiment analysis, treated in isolation.

In this thesis, we contribute to the understanding of extra-propositional meaning in natural language understanding, by providing a comprehensive account of the semantic phenomena that occur beyond simple assertions and examining how a coherent discourse is formed from lower level semantic elements. Our approach is linguistically based, and we propose a general, unified treatment of the semantic phenomena involved, within a computationally viable framework. We identify semantic embedding as the core notion involved in expressing extra-propositional meaning. The embedding framework is based on the structural distinction between embedding and atomic predications, the former corresponding to extra-propositional aspects of meaning. It incorporates the notions of predication source, modality scale, and scope. We develop an embedding categorization scheme and a dictionary based on it, which provide the necessary means to interpret extra-propositional meaning with a compositional semantic interpretation methodology. Our syntax-driven methodology exploits syntactic dependencies to construct a semantic embedding graph of a document. Traversing the graph in a bottom-up manner guided by compositional operations, we construct predications corresponding to extra-propositional semantic content, which form the basis for addressing practical tasks. We focus on text from two distinct domains: news articles from the Wall Street Journal, and scientific articles focusing on molecular biology. Adopting a task-based evaluation strategy, we consider the easy adaptability of the core framework to practical tasks that involve some extra-propositional aspect as a measure of its success. The computational tasks we consider include hedge/uncertainty detection, scope resolution, negation detection, biological event extraction, and attribution resolution. Our competitive results in these tasks demonstrate the viability of our proposal.