| This thesis links a linguistic phenomenon, reported speech, within a certain context, newspaper articles, to a pragmatic interpretation, namely that the complement clause contains the important primary information and that the role of the matrix clause is to provide evidence for this primary information. This is captured in the evidential analysis proposed in this thesis. The evaluative context of the matrix clause projects up from the lexical semantics of the two major components, namely the reporting verb and the source (the subject of the reporting clause) and combines with the context of the surrounding text.; This thesis builds on current trends in computational linguistics, taking advantage of large on-line corpora and developing a lexical semantics. It goes beyond the main focus of current research to automatically extract lexical data, suggesting that a theoretically guided, semi-automatic corpus analysis can yield unexpected, (i.e. not automatically detectable) results that can influence the theoretical basis of a lexical semantics.; This thesis introduces a lexical semantics for reporting verbs and some frequent source descriptions within the framework of Pustejovsky's Generative Lexicon theory (Pustejovsky, 1991). The thesis goes, however, beyond lexical semantics, connecting the lexical properties with a text representation scheme, MTR, which is designed to make explicit the semantic (and pragmatic) information implicit in the reported speech sentence. Moreover, MTR offers a new way to represent texts introducing a device called profile structure, that groups reported speech sentences according to the source. Profiles complement traditional text representation devices such as coherence structure of sentences and temporal structure of events.; This thesis thus spans from lexical semantics to text analysis, suggesting how to integrate current methods and novel devices for a concrete task. This thesis attempts to not only provide new insights into the analysis of reported speech and newspaper articles in general, but also to contribute to the design of more robust NL systems that can analyze real data reliably. |
