Current text generation systems are not very good at choosing lexical items that signal a coherence relation, such as weil, deshalb, nachdem (so-called discourse markers). Instead, they either adhere to the "one-relation-one-marker" strategy, or consider only a small set of discourse markers. Yet, language offers many discourse markers to express the same coherence relation. These markers signal pragmatic and semantic aspects on top of the coherence relation itself. Not paying attention to these fine-grained differences in meaning severely restricts the expressiveness of text generators, and may also decrease the readability or even endanger the correct interpretation of a text. Choosing the appropriate discourse marker in a given context requires detailed knowledge of the function and form of a wide range of discourse markers, and a generation architecture that integrates discourse marker choice into the overall generation process. Both aspects are not solved in a satisfactory manner by available descriptive studies of discourse markers and by current text generation systems. This thesis addresses these two issues and makes contributions to the two distinct areas of linguistic description and discourse marker choice in multilingual text generation.

The thesis starts with a survey of current research on discourse marker representation, and choice and on approaches to discourse marker representation in multilingual text generation, and discusses methodological questions, in particular the question of how to arrive at the set of lexical items functioning as discourse markers. Here, the thesis proposes a "Test for discourse markers in German texts" which provides criteria for identifying discourse markers in text. Finally, the focus of the research – German temporal discourse markers – is motivated; English temporal markers are considered to demonstrate the multilingual perspective.

The second part addresses issues of discourse marker analysis and representation:

First, those properties that have to be accounted for in a comprehensive study of German temporal discourse markers are determined. Based on this, fine-grained descriptions of the semantic, pragmatic, and syntactic characteristics of 27 German temporal marker are given. The results of the analysis are then merged in a single functional representation, i.e. a functional classification of German temporal conjunctive relations. The proposed networks are then compared to existing accounts in English and Dutch to provide a basis for a multilingual representation.

Discourse markers are generally said to signal coherence relations. Hence, these relations are the major constraints in discourse marker choice. This thesis argues that the "atomic" relation definitions used in Rhetorical Structure Theory – an approach to discourse representation that is very popular in text generation - cannot adequately describe the coherence relations holding between discourse segments in text and do not meet the demands of discourse marker representation and choice. It is demonstrated that the contributions from three dimensions (ideational, interpersonal, textual) need to be distinguished. This thesis proposes a paradigmatic description of coherence relations along these three dimensions, yielding composite relations as well as a methodology to derive such a description from a text corpus.

The final part of the thesis addresses the issues of lexical modelling and discourse marker selection in (multi-lingual) text generation. The thesis argues that – given the assumption that discourse marker choice is one task among others in the sentence planning phase – a discourse marker lexicon is the most suitable representation for discourse marker knowledge to support the motivated choice of discourse markers in (multilingual) text generation systems. A discourse marker lexicon is introduced as a generic resource for storing discourse marker meaning and usage. The global organisation of the lexicon and the shape of the individual lexicon entries are presented, following the recommendations of the Expert advisory group on language engineering standards (EAGLES) for lexical semantic encoding, and sample entries for German and English temporal discourse markers are given. Then, a generation architecture is proposed that exploits the discourse marker lexicon at sentence planning level and a computational model for discourse marker choice is outlined. Several (bilingual) examples illustrate how the choice procedure – deploying the discourse marker lexicon as major resource - produces alternative verbalisations of temporal coherence relations holding between two situations.