Coherence in Spoken Discourse

Heike Tappe, Frank SchilderCOLING 1998

TL;DR

This paper explores how a discourse grammar, specifically a Tree Description Grammar (TDG), can be adapted to analyze spontaneous spoken language, which is often less structured than written text. By analyzing subjects describing sketch-map creations in real-time, the authors demonstrate that TDG can capture the dynamic planning processes and less globally coherent structures inherent in spoken discourse, extending beyond descriptive theories.

What Problem Does This Paper Address?

Analyzing spontaneous spoken discourse presents a significant challenge for existing discourse grammars. Many established theories, like SDRT and RST, are primarily designed for written text, which is typically produced through a deliberate planning process, resulting in a more globally coherent structure. Spoken discourse, however, is often characterized by on-the-fly planning, hesitations, and less predictable connections between utterances. This paper investigates whether a flexible discourse grammar, based on Tree Description Grammars (TDG), can effectively model the dynamics and inherent incoherence of spoken language, particularly when speakers are simultaneously conceptualizing, planning, and producing their speech.

Methods

The authors employ a Tree Description Grammar (TDG) framework, extended with feature structures to encode semantic content and discourse prominence. This approach allows for the representation of the "dynamics" of discourse structure, going beyond purely descriptive accounts. Specifically, they introduce features like CONT(ENT) for sentence meaning and PROMI(NENT) to capture the notion of "openness" in discourse, reflecting how utterances are subordinated by others. The TDG's ability to handle dominance and parent relations between nodes in a tree structure is leveraged to model how discourse segments are attached and integrated. For spoken discourse, they propose incorporating new rhetorical relations that specifically describe planning processes, such as plan_comment.

Datasets

The study uses a dataset of online verbalizations of subjects describing the pixel-per-pixel development of sketch-maps on a computer screen. These sketch-maps were created by students depicting a route between two university buildings. The verbalizations were elicited in an experimental setting where new subjects watched replays of these drawings and described them in real-time as the graphical objects appeared. This method aims to simulate the pressure of simultaneous conceptualization, planning, and production.

Key Insights and Ideas

  • Spoken Discourse is Less Globally Coherent: The analysis reveals that spoken discourse, especially when produced under real-time pressure, exhibits less global coherence than written text.
  • TDG Can Model Spoken Discourse Dynamics: The proposed flexible discourse grammar based on TDG, with its ability to encode dynamics and planning processes, proves capable of analyzing spontaneous speech.
  • Planning Processes Surface in Discourse: The study highlights how speakers' internal planning processes (e.g., searching for words, formulating expectations) can manifest as rhetorical questions or meta-comments within the discourse itself.
  • New Rhetorical Relations are Needed: To adequately model spoken discourse, new rhetorical relations, such as plan_comment, are introduced to capture these planning-related utterances.
  • Right Frontier Principle Holds: Despite the less coherent nature, the "right frontier principle" (where new discourse segments attach to the most recently opened nodes) appears to be maintained, even when digressions occur.

Key Results

The paper demonstrates the formalization of a discourse structure for the first three utterances of the example text, showing an elaboration relation between U1 and U2, and a narration relation between U2 and U3. It also illustrates how utterances like U4-U6, which express planning and searching processes, can be integrated into the discourse structure using new rhetorical relations like plan_comment. The analysis shows that even utterances that seem tangential, like rhetorical questions, can be incorporated, and the discourse structure can be extended to accommodate them. The paper presents a visual representation (Figure 4) of the discourse structure that includes these planning elements.

Limitations and Future Work

The authors acknowledge that their approach extends existing discourse grammars to handle the complexities of spoken language. Future work will focus on further refining the interaction between planning sequences, discourse structure, and intentional structure. The paper also notes that sequences involving rhetorical questions describing planning processes would likely not be found in written text, implying a distinction in how these phenomena are handled across modalities.

Summary updated

BibTeX
@inproceedings{tappe1998coherence,
  author = {Heike Tappe and Frank Schilder},
  title = {{Coherence in Spoken Discourse}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1998},
}
Figure