Projection in discourse:A data-driven formal semantic analysis

A sentence like "Bertrand, a famous linguist, wrote a book" containsdifferent contributions: there is a person named "Bertrand", he isa famous linguist, and he wrote a book. These contributions convey differenttypes of information; while the existence of Bertrand is presented as giveninformation---it is presupposed---the other contributions signal newinformation. Moreover, the contributions are affected differently bylinguistic constructions. The inference that Bertrand wrote a bookdisappears when the sentence is negated or turned into interrogative form,while the other contributions survive; this is called 'projection'. In thisthesis, I investigate the relation between different types of contributionsin a sentence from a theoretical and empirical perspective. I focus onprojection phenomena, which include presuppositions ('Bertrand exists' inthe aforementioned example) and conventional implicatures ('Bertrand isa famous linguist'). I argue that the differences between the contributionscan be explained in terms of information status, which describes how contentrelates to the unfolding discourse context. Based on this analysis, I extendthe widely used formal representational system Discourse RepresentationTheory (DRT) with an explicit representation of the different contributionsmade by projection phenomena; this extension is called 'Projective DiscourseRepresentation Theory' (PDRT). I present a data-drivencomputational analysis based on data from the Groningen Meaning Bank,a corpus of semantically annotated texts. This analysis shows how PDRT canbe used to learn more about different kinds of projection behaviour.These results can be used to improve linguistically oriented computationalapplications such as automatic translation systems