Topic modeling is an effective technique for uncovering hidden themes in documents and is widely used in text mining and data analysis. Recently, large language model (LLM) based topic models generate topics via prompts and assign them to documents, yielding more natural and readable topics. However, the conventional assignment approach has drawbacks: it cannot produce a topic distribution for a document, may generate topics that are too broad or too narrow, and its resource consumption grows linearly with the number and length of documents. These issues are critical in large‑scale industrial settings that demand high‑quality, in‑depth analysis. To address this, we introduce the SeLATM framework, which performs segment‑level topic generation and refines topics through agentic feedback loops. Experiments on several datasets show that SeLATM maintains or improves performance while substantially reducing LLM resource usage.
Review