| Entities realized in syntactically prominent positions are preferred antecedents for pronominal reference. This has been demonstrated in numerous psycholinguistic experiments (e.g., Hudson-D'Zmura and Tanenhaus, 1997; Mathews and Chodorow, 1988) and corpus investigations (Arnold, 1998b). However, for many verbs in English---especially eventive verbs---syntactic role and semantic role are often conflated. That is, syntactic subjects are often semantic agents and carry more PROTO-agent entailments (e.g., sentience, volition; see Dowty, 1991), while syntactic objects are often semantic patients and carry more PROTO-patient entailments (e.g., undergo change-of-state, causally-affected). Thus, it could be said that entities realized in prominent roles on a hierarchy of semantic roles (e.g., agent > patient as in Jackendoff, 1972) are preferred antecedents for pronominal reference. As such, the central question of this dissertation can be worded as follows: Is it the case that what has heretofore been observed as a result of syntactic prominence is in fact a result of semantic prominence?; I take as a starting point a generalized model of discourse processing in which the relative salience of discourse referents in the current context is seen as influencing subsequent reference to those referents as well as the form of referring expression (cf., Grosz et al., 1995; Kintsch and van Dijk, 1978) and then incorporate a detailed notion of discourse salience in which two prominence factors contribute to overall salience: syntactic and semantic prominence. Using this model as a foundation, I compare the relative effects of syntactic and semantic prominence in two investigative paradigms: psycholinguistic experimentation and corpus analysis. Results from both investigations give evidence that both syntactic and semantic prominence contribute to the salience of discourse referents. These results are analyzed with respect to two approaches to determining semantic prominence---a frame semantic (Fillmore, 1968, 1976) approach using the FrameNet system (Baker et al., 1998) and a PROTO-role approach using the PROTO-role entailments of Dowty (1991). A comparison of these two approaches and implications of the main findings for psycholinguistic models and computational implementations of discourse processing are discussed. |
