一句话：提出 DIGEST 框架，利用层次化主题引导提升零样本 LLM 对多主题支持工单对话的摘要质量。
云服务支持工单对话通常冗长、碎片化，且多个主题交织演进，造成信息过载、延迟问题解决并可能掩盖关键细节。监督式或主题感知摘要方法虽能提升连贯性，但依赖大量标注数据与重新训练，成本高且难以适应动态环境。零样本大语言模型（LLM）缓解了这些限制，但直接用于多方事件对话仍受重叠发言、焦点转移和大量噪声的挑战。为此，作者提出 DIGEST，一种面向主题的摘要框架，将层次化主题引导引入零样本 LLM 摘要。DIGEST 先进行自顶向下分割，将粗粒度线程拆解并细化为子主题，再通过自底向上聚合，把局部摘要整合为连贯叙述，同时保持各问题之间的独立性。实验结果表明，DIGEST 优于基于 LLM 的摘要基线。
Ticket-driven dialogues in cloud service systems, such as those in ByteDance's Volcano Engine, are typically lengthy and fragmented. They are often interleaved across multiple evolving topics, which creates information overload, delays resolution, and risks obscuring critical details. While supervised or topic-aware summarization approaches can improve coherence, they require large annotated datasets and retraining, which are costly and poorly suited to dynamic environments. Zero-shot large language models (LLMs) alleviate these constraints, yet direct application to multi-party incident dialogues remains challenging due to overlapping voices, shifting focus, and substantial noise. To address this, we propose DIGEST, a topic-wise summarization framework that introduces hierarchical topic guidance into zero-shot LLM summarization. DIGEST first performs top-down segmentation to disentangle coarse threads and refine them into subtopics, then applies bottom-up aggregation to integrate local summaries into coherent narratives while preserving independence across issues. Experimental results demonstrate that it outperforms LLM-based summarization baselines.
@inproceedings{ChenLHZWW26,
title = {Topic-wise Summarization of Support Ticket Dialogue via LLM},
author = {XiaoLei Chen and Fengrui Liu and Xiao He and Tieying Zhang and Peng Wang and Wei Wang},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805250}
}