Encoding Discourse Structure: Comparison of RST and QUD

What it is about

The authors annotate 28 German texts (blog posts and the podcast transcripts they describe) with both RST trees and QUD trees, convert each RST tree mechanically into QUD-tree form (a relation node becomes a question node over its children, nuclearity dropped), and compare the unlabelled branching with a PARSEVAL-style score: does a node with the same yield exist in the other tree, in both directions, harmonic mean. Average similarity is 0.74, comparable to agreement between annotators within one framework, and higher for monologue blogs than for dialogue. The qualitative comparison shows what QUD trees lose: they do not mark which segment is central, and concession and contrast can only be modelled with an extra subquestion that hides which part is the expectation and which the violation; restatements and summaries also lose their function. QUD trees, in turn, capture topic progression and speaker changes in dialogue more naturally than RST.

Ideas

• I1. Use unlabelled yield-matching (PARSEVAL over spans) to compare any two discourse structures Mova builds for the same text: the QUD stack of world/src/qud.rs against the grouping implied by the frame links of prag/src/frames.rs, or against a teacher-made tree. It needs no agreement on labels or question wording; implement it once in Rust next to the frame bench scorer (prag/src/fbench.rs).
• I2. Keep contrast and concession as explicit links on frames rather than leaving them to the QUD stack, since QUD trees cannot say which side is the expectation: a but/although/yet clause gives a contrast link with the expected side marked. «The crow wanted to keep the cheese, but she opened her beak» then supports questions like «What did the crow do although she wanted to keep the cheese?». Check how many course questions turn on but/although.
• I3. Where Mova needs centrality (choosing the main event of a paragraph for a summary or for «what is the story about?»), take it from frame links, not from the QUD stack: the frame that others attach to as cause or elaboration is central.

For Mova

36. Discourse relations and coherence §3 (RST), §9 and §11 (QUD stack, links between frames, not yet). Code: world/src/qud.rs, prag/src/frames.rs (contrast link), prag/src/fbench.rs (span-matching scorer). Measures: yield-matching similarity between structures; course questions with but/although.

Topics: Discourse analysis

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.codi-1.11

@inproceedings{shahmohammadi2023encoding,
  title = {{Encoding Discourse Structure: Comparison of RST and QUD}},
  author = {Shahmohammadi, Sara and Seemann, Hannah and Stede, Manfred and Scheffler, Tatjana},
  booktitle = {Proceedings of the 4th Workshop on Computational Approaches to Discourse (CODI 2023)},
  year = {2023},
  doi = {10.18653/v1/2023.codi-1.11},
  url = {https://doi.org/10.18653/v1/2023.codi-1.11}
}

References

4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Lynn Carlson and Daniel Marcu. 2001. Discourse tagging reference manual. ISI Technical Report ISI-TR-545, 54(2001):56. find
2. Jonathan Ginzburg. 1996. Dynamics and the semantics of dialogue. Logic, language and computation, 1:221–237. find
3. Julie Hunter and Márta Abrusán. 2017. Rhetorical structure and QUDs. In New Frontiers in Artificial Intelligence, pages 41–57. Springer International Publishing. find
4. Wolfgang Klein and Christiane von Stutterheim. 1987. Quaestio und referentielle Bewegung in Erzählungen. Linguistische Berichte, 109:163–183. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/shahmohammadi-2023-encoding-discourse-structure-comparison-rst/