Compositionality and the semantics of nominals

In this dissertation, the semantic and syntactic properties of the set of nouns which are known as agentave nominals are analyzed in the context of current concerns in computational lexical semantics. The aim of the analysis is twofold. From a theoretical perspective, it is argued that a rich enough view of lexical structure explains the polysemous behavior of agentive nouns and provides the basis for systematically capturing the semantic relation that a derived nominal bears to a correlated verbal form. From a computational perspective, it responds to the need of creating more efficient and linguistically motivated lexicons for natural language processing.;Adopting the theoretical framework of Generative Lexicon theory, it is shown that, by recasting the notion of agentive nominalizations from the perspective of lexical semantics rather than morphology, the semantic interpretation of derived nominal forms is systematic and predictable. This is achieved by viewing morphological operations as instances of creative use of language that consistently derive new forms which are independently motivated in the lexicon. Thus, the correlations between lexically encoded base forms and morphologically derived forms provide a new paradigm of data showing that the distinction between different types of agentive nominals can be expressed in terms of stage-level and individual-level properties in the domain of nominals. Focusing on the role of events in the semantics of nouns, it is shown that stage-level and individual-level agentive nouns differ in the type and the quantification of their defining event. This is rejected in a variety of syntactic and semantic properties including temporal and locative modification, binding of pronouns, genericity, temporal constraints with the matrix predicate, among others. The lexical semantic account presented here strongly motivates the interest of adopting a generative lexical model of representation, which, not only makes the desired distinctions, but can be exploited in a variety of natural language processing tasks requiring linguistically-motivated solutions.