Open Collections

Ivanova, Inna

University of British Columbia

This thesis introduces a discourse-guided approach for generating text from semi-structured data - knowledge graphs and image scene graphs. We provide a novel architecture that integrates discourse planning as an intermediary structuring step, with the objective of enhancing coherence, readability, and the overall quality of generated content. The proposed method reorders input graph nodes into a coherent discourse sequence prior to decoding, utilizing both Pointer Networks and Large Language Models (LLMs) to represent discourse structures. Our experiments focus on two distinct datasets - Agenda (scientific abstracts) and Visual Genome (image captioning) - illustrating that explicit discourse planning consistently enhances performance across standard natural language generation metrics and improves output quality as evaluated by both human and LLM-based assessments. This thesis presents a generalizable approach for integrating discourse structure into neural text generation systems and emphasizes the potential of large language models as both planners and evaluators in natural language generation tasks.

Text

2025-07-04

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/91497

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/