基於大型語言模型的自動化主題標示研究

A Study of Automated Topic Labeling Based on Large Language Models

本研究提出一種基於大型語言模型（LLM）的自動化主題標示方法，能夠為主題模型中的每個主題產生有意義的詞語和摘要標籤，並提出兩種生成策略：簡要描述語標示（CDL）和上下文增強標示（CEL）。除了對生成結果進行質性觀察和傳統的穩定性和主題相關性測量外，還以主題指派任務評估涵蓋性和區別性。實驗結果顯示，CDL傾向產生簡明的學科術語，具有高穩定性並相當接近描述語的語意；CEL則常生成專業術語，儘管多次生成的詞彙有些微變化，語意仍保有相當高的一致性。雖然CEL在涵蓋性與區分性上表現較優，但兩種策略的結果都可接受且具互補性，並可根據應用情境選擇適合策略。未來研究建議著重生成策略的進一步優化和整合，增強主題模型的理解與應用。

This study proposes an automated topic labeling method based on large language models (LLMs), capable of generating meaningful term and summary labels for each topic within a topic model. Two generation strategies are introduced: Concise Descriptor Labeling (CDL) and Context- Enhanced Labeling (CEL). In addition to qualitative observation and conventional measures of stability and topical relevance, this study further evaluates coverage and discriminativeness through a topic assignment task. The experimental results show that CDL tends to produce concise and general disciplinary terms, characterized by high stability and close semantic alignment with the topic descriptors. In contrast, CEL often generates specialized technical terminology; although lexical variations may occur across multiple generations, semantic consistency remains high. Though CEL demonstrates slightly better performance in terms of coverage and discriminativeness, both strategies are acceptable and complementary, allowing researchers to select the appropriate approach depending on the interpretive and application context of the topic model. Future research may focus on refining these generation strategies and exploring their integration to enhance the applicability and interpretability of topic models.