TL;DR
This paper compares two prominent theories of discourse structure, Rhetorical Structure Theory (RST) and Questions Under Discussion (QUD), by applying them to a novel corpus of German blog posts and podcast transcripts. The study finds significant overlap in the structural information captured by both frameworks, suggesting they can be complementary, but also highlights distinct strengths and weaknesses, particularly in representing discourse centrality and dialogic features.
What Problem Does This Paper Address?
Understanding how discourse is structured is fundamental to natural language processing, from text summarization to dialogue systems. Two dominant theoretical frameworks exist: Rhetorical Structure Theory (RST), which focuses on the intentional and rhetorical relationships between text segments, and the Questions Under Discussion (QUD) model, which frames discourse as a series of questions and answers, emphasizing information structure. While both aim to represent discourse as a tree, they originate from different fields and focus on different phenomena. Researchers have explored applying one theory to the other's domain, but a systematic, empirical comparison on a diverse, multi-modal corpus was lacking. This paper addresses this gap by providing a direct, quantitative and qualitative comparison of RST and QUD annotations on the same texts, aiming to understand their correspondences and divergences.
Methods
The core of this research involves a parallel annotation of texts using both RST and QUD frameworks. To enable quantitative comparison, the authors developed a method to convert RST trees into a QUD-like format. This conversion treats RST discourse relations as implicit questions in the QUD structure, with the satellite and nucleus of the relation becoming children of this implicit question node. The similarity between the converted RST trees and the manually annotated QUD trees was then evaluated using a variant of the PARSEVAL measure, adapted for unlabelled trees. This measure assesses the overlap in tree structures based on the "yield" (the text segments dominated by a node). Beyond quantitative analysis, a qualitative comparison was conducted to identify specific areas where the frameworks align or diverge, focusing on how complex discourse units are handled, the translatability of specific discourse relations, and the representation of dialogic features like speaker changes.
Datasets
The study utilizes a novel multi-media corpus of 28 German texts, comprising 14 blog posts (monologue) and 14 podcast transcripts (dialog). These texts are related, with blog posts often serving as descriptions of what is discussed in the podcasts. This corpus is unique in that it is annotated in parallel with both RST and QUD structures, and the authors make the annotated corpus available for future research. The blog posts averaged 26 EDUs (Elementary Discourse Units), while the podcast transcript chunks averaged 17 EDUs.
Key Insights and Ideas
- Complementary Strengths: While both RST and QUD capture significant structural information, they highlight different aspects of discourse. QUD excels at representing topic progression and speaker changes in dialog, which can be less explicit in RST. Conversely, RST provides a more nuanced representation of intentional and rhetorical relationships.
- Centrality Matters: A key finding is that QUD structures, by their nature, do not explicitly indicate the "centrality" of discourse segments in the same way RST does through its nucleus-satellite distinction. This means that while QUD can group segments under a common question, it doesn't inherently prioritize one segment as the main point.
- Relation Translatability: Some RST relations, like "concession" and "contrast," are challenging to directly map to QUD. While QUD can represent these through nested sub-questions, it loses the explicit indication of expectation/violation or the direct contrastive meaning present in RST.
- Dialog vs. Monologue: The similarity between RST and QUD annotations was higher for monologue texts (blog posts) than for dialog (podcast transcripts). This suggests that the agreement on how to represent the overall discourse progression is more straightforward when there's a single author's intent, compared to the negotiation of topics and turns in a conversation.
- RST's Granularity: RST tends to build more complex discourse units, sometimes leading to mismatches with QUD's more local segmentation, especially when information-structural factors suggest a discourse contribution.
Key Results
The quantitative analysis revealed a substantial overlap in the discourse structures identified by RST and QUD, with an average similarity (micro-averaged parseval score) of 0.74. This score is comparable to inter-annotator agreement within a single framework, indicating a high degree of correspondence. Blog posts showed a higher similarity score (0.76) than podcast transcripts (0.73). Qualitatively, the study found that while complex discourse units often align, QUD trees do not indicate segment centrality, and RST trees struggle to easily represent features like speaker changes in dialog, which QUD handles more naturally.
Limitations and Future Work
The authors acknowledge several limitations. Firstly, the annotations were performed by distinct researchers, and while double-checked, alternative interpretations are always possible in discourse analysis. Secondly, the study lacks direct access to the discourse creators' intentions, which is a common limitation in corpus studies. For future work, the authors plan to add a second annotator for each framework to calculate inter-annotator agreement and will make the annotated corpus and conversion tool publicly available to facilitate further research. They also suggest that future work could explore evaluating the equivalence of natural language questions in QUD trees, a task not undertaken here.
Summary updated