Official code for ACL 2025 paper: Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation.
This is the overview of Def-DTS framework.
-
We introduce LLM reasoning techniques to DTS for the first time, consolidating insights from previous methodologies into a coherent and deductive prompt design.
-
We reformulate DTS as an utterance-level intent classification task by implementing intent classification as a core component of a multi-step reasoning process, enabling flexible and task-agnostic prompting.
-
Our method empirically demonstrates superior performance across nearly all comparative baselines, underscoring the efficacy of prompt engineering in DTS.
-
Through an in-depth analysis of our approach's reasoning results, we shed light on the challenges LLM reasoning faces in DTS and discuss the possibility of using LLM as a DTS auto-labeler.
Please refer to the paper for details.
@misc{lee2025defdts,
title={Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation},
author={Seungmin Lee and Yongsang Yoo and Minhwa Jung and Min Song},
year={2025},
eprint={2505.21033},
archivePrefix={arXiv},
primaryClass={cs.CL}
}