Annotation Science, a discipline dedicated to developing and maturing methodology for the annotation of language resources, is playing a prominent role in the fields of computational and corpus linguistics. While progress in the search for the right annotation model and format is undeniable, these results only sparsely become manifest in actual solutions (i.e. software tools) that could be used by researchers wishing to annotate their resources right away, even less so for resources of spoken language transcriptions. The dissertation presents a solution consisting of a data model and an annotation tool that tries to fill this gap that exists between "annotation science" and the practice of transcribing spoken language in the area of discourse analysis and pragmatics, where the lack of ready-to-use annotation solutions is especially remarkable.
Transcriptions of discourse vary from other language resources in many ways: they depend on the temporal organization of the speaker's contributions. To analyze phenomena like turn-taking, it is crucial to exactly record what happens when. In discourse, there is not one primary data like in most written texts: every speaker's utterances make up for one stream of primary data, and these streams of primary data can and frequently do overlap. Departing from these and other observations, a data model for the transcription of spoken discourse has to: supply a timeline to represent the temporal properties of the speaker's utterances, allow for overlap of speaker's utterances and overlap of annotations, be prepared for different options for the segmentation of the transcribed material and support most existing writing systems. Existing data models are distributed on an axis between two poles. The one pole consists of a model where language is seen as a hierarchical structure in which the down most layer represents the actual textual content (OHCO, Ordered Hierarchy of Content Objects) (3). These models always have to introduce special "workarounds" to deal with overlapping hierarchies or speaker contributions. The other pole is represented by a model that is based on directed, acyclic graphs and is centered on the aspect of temporal relations of the respective units.
Transcription graphs, on which the EXMARaLDA transcription format and toolset is built, specify and substantiate the annotation graph formalism. The EXMARaLDA data format, in particular, introduces the possibility to segment the textual content of the transcriptions according to transcription conventions. Departing from a so-called "basic transcription", that in most cases only contains segments motivated by changes in the constellation of the discourse (speaker changes, interruptions) that are directly linked to the explicit timeline, the EXMARaLDA system allows for the automatic segmentation into linguistically motivated segments (like words, utterances, sentences, turns) that are defined through rules laid down in the transcription conventions and do not necessarily link to distinct points on the timeline. While it is possible to add annotations to the former type of segments by simply adding additional annotation tiers to the transcription in the transcription tool, there is no mechanism, neither in software nor in the data format, to add annotations to the latter.
Based on these premises, an annotation format and software tool were developed to facilitate annotation of these segments. Since tools and large and valuable corpora for the EXMARaLDA format already exist, one indispensable premise was that no changes on the tools and the existing format should be necessary. This precondition determined the usage of stand-off annotation: stand-off annotation stores annotations and the annotated content in different locations (i.e. files), and connect them through pointers.
Annotations themselves are modeled as feature structures, following the TEI's recommendations for feature structures that are also an ISO standard. Feature structures model information as attribute/value pairs, where the value can either be atomic or another attribute/value pair. That way, they allow for the easy modeling of simple attribute/value combinations, but also allow for much more complex annotation structures like trees. Furthermore, feature structures offer an established method of creating libraries of frequently used features that can be utilized by pointing at them.