Propositionaler und Non-at-issue Inhalt für die Textgenerierung: Eine Untersuchung zum QUD-Ansatz für die Diskursstrukturierung

• Antragsteller: Dr. Anton Benz; Professor Dr. Ralf Klabunde
• Fachliche Zuordnung: Allgemeine und Vergleichende Sprachwissenschaft, Experimentelle Linguistik, Typologie, Außereuropäische Sprachen
• Förderung: Förderung von 2019 bis 2024
• Projektkennung: Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 427866300

• Erstellungsjahr: 2023

Zusammenfassung der Projektergebnisse

Question under discussion (QUD) based approaches to text understanding assume that texts can be analyzed as complex, structured answers to implicit (or explicit) questions, the QUDs. This approach turned out to be fruitful in explaining a wide range of pragmatic phenomena, from dialogue moves to information structure to temporal progression. In this project, we inverted the perspective. Instead of the usual approach of reconstructing the QUD structure for a given text, we aimed at generating a text, given a QUD structure. Changing this perspective raised several new theoretical, empirical and computationally motivated research questions that we addressed in this project. Together with a suitable annotation tool, we created a QUD-based annotated corpus of driving reports which provided the data basis for the clarification of theoretical issues and the development and evaluation of the generation system. We developed new insights on the at-issue/non-at-issue distinction as well as on the relation between rhetorical relations and QUDs. We furthermore showed how evaluative expressions – typically not addressed by a QUD – can be integrated into a QUD-oriented Natural Language Generation System, and how the use of such expressions can be motivated by regression methods.

Projektbezogene Publikationen (Auswahl)

• Annotating QUDs for generating pragmatically rich texts. In: Proceedings of the Workshop on Discourse Theories for Text Planning. Dublin, Ireland, December 2020. Association for Computational Linguistics; 10–16.
  Christoph Hesse; Anton Benz; Maurice Langner; Felix Theodor & Ralf Klabunde
  
• Discrepancies Between Database- and Pragmatically Driven NLG: Insights from QUD-Based Annotations. In: Dagmar Gromann, Gilles Sérasset, Thierry Declerck, John P. McCrae, Jorge Gracia, Julia Bosque-Gil, Fernando Bobillo, and Barbara Heinisch (eds.) 3rd Conference on Language, Data and Knowledge (LDK 2021), volume 93 of Open Access Series in Informatics (OASIcs), pages 32:1–32:9, Dagstuhl, Germany, 2021. Schloss Dagstuhl – Leibniz-Zentrum für Informatik.
  Christoph Hesse; Maurice Langner; Anton Benz & Ralf Klabunde
  
• Modeling argumentative contrast structures using QUD-trees. In: International Workshop on the Expression of Contrast and the Annotation of Information Structure in Corpora. Leuven, Belgium, 18-19 November 2021.
  Christoph Hesse; Anton Benz & Ralf Klabunde
  
• The gap between QUD-based topic determination and learning-based topic extraction for NLG. In: Meaning in Context: Pragmatic Communication in Humans and Machines. Workshop at the 35th Conference on Neural Information Processing Systems (NeurIPS 2021), December 13.
  Maurice Langner & Ralf Klabunde
  
• Linguistik im Sprachvergleich. Springer Berlin Heidelberg.
  Ralf Klabunde; Wiltrud Mihatsch & Stefanie Dipper (eds.)
  
• QUDs and discourse relations: Non-at-issue information in texts. In: Discourse studies and linguistic data science (DisLiDas), Jerusalem, Israel, 24 May 2022.
  Christoph Hesse; Ralf Klabunde; Anton Benz & Maurice Langner
  
• Realizing a Denial of Expectation in Pipelined Neural Data-To-Text Generation. In: Proceedings of the 6th Workshop on Advances in Argumentation in Artificial Intelligence. Udine, Italy, November 2022.
  Maurice Langner & Ralf Klabunde
  
• Testing focus and non-at-issue frameworks with a question-under-discussion-annotated corpus. In: Proceedings of the Language Resources and Evaluation Conference. Marseille, France, 5212–5219.
  Christoph Hesse; Maurice Langner; Ralf Klabunde & Anton Benz
  
• Validating Predictive Models Of Evaluative Language For Controllable Data2Text Generation. In: Proceedings of the 16th International Natural Language Generation Conference. Prague, Czechia, September 2023.
  Maurice Lagnner & Ralf Klabunde
  
• 9 Contrast, concession, and QUD-trees. On the Role of Contrast in Information Structure, 225-262. De Gruyter.
  Hesse, Christoph; Klabunde, Ralf & Benz, Anton
  
• Controllability of evaluative language generation in neural data-to-text NLG. Dissertation, completed in January 2024.
  Maurice Langner