LLM-Summarize-BERTopic

Topic Modeling Enhancement using Summaries Generated by LLM Models

Published in the Proceedings of the International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA 2025), Antalya, Turkey, August 2025. IEEE Xplore.

• Paper: https://doi.org/10.1109/acdsa65407.2025.11166276
• Authors: Haya El Ghalayini, Min Liu, Bhagyesh Patel, Hetkumar Patel, Rasheed Amanzai (Sheridan College)

The idea

Topic models struggle with long, unstructured documents: topics overlap, and long texts have to be chopped up before a model like BERTopic can use them. This research asks whether an LLM can do that preparation better.

Two approaches compared

	Method 1: paragraphs	Method 2: LLM summaries
prepare	split each document into paragraphs	an LLM splits the document into token segments and summarizes each one
model	BERTopic on the paragraphs	BERTopic on the summaries

Result

Summarizing the documents with an LLM first, then running BERTopic on the summaries, performed better than running BERTopic on paragraphs. See the paper for the datasets, metrics and full results.

Citation

H. El Ghalayini, M. Liu, B. Patel, H. Patel and R. Amanzai,
"Topic Modeling Enhancement using Summaries Generated by LLM Models,"
Proc. International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA 2025), Antalya, Turkey, 2025.
doi: 10.1109/ACDSA65407.2025.11166276

Contact: Bhagyesh Patel