Text Mining in the Era of Large Language Models
Resumo
Large Language Models (LLMs) are increasingly adopted in individual Text Mining (TM) tasks, but their role across complete analytical workflows remains insufficiently characterized. This dissertation investigates LLMs as a transversal enhancement layer across a five-stage TM process, covering text representation, weak supervision, clustering, and agentic orchestration. Controlled experiments show that dense embeddings outperform lexical representations; zero-shot LLM annotation reaches macro-F1 0.756 versus 0.824 under full supervision, while distillation into a lightweight classifier reaches 0.706; consensus pseudo-labels reduce the gap between simple and deep clustering pipelines; and LLM agents reliably orchestrate classification workflows. Applications in public policy, mental health, education, and risk management demonstrate the broader social and practical impact of the findings.
Referências
Paulo Mann, Matheus Yasuo Ribeiro Utino, Elton Hiroshi Matsushima, and Aline Paes. 2025. The Topics of Depression on Social Networking Sites. Journal of the Brazilian Computer Society 31, 1 (Oct. 2025), 771–806.
Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. MTEB: Massive Text Embedding Benchmark. In EACL. ACL, Dubrovnik, Croatia, 2014–2037. DOI: 10.18653/v1/2023.eacl-main.148
Matheus Yasuo Ribeiro Utino, Vin’ıcius Teixeira Carvalho, and RicardoMarcondes Marcacini. 2026. Learning risk matrix from textual data through graph neural networks. Humanities and Social Sciences Communications (2026).
Matheus Utino, Marcos Gôlo, Marcelo Turine, and Ricardo Marcacini. 2026. Lightweight Adherence Prediction and Justification for Government Strategic Priorities. In LASDIGOV (Gramado/RS). SBC, Porto Alegre, RS, Brasil, 217–228. DOI: 10.5753/lasdigov.2026.23817
Matheus Utino and Paulo Mann. 2025. From Zero-shot to Self-generated References: Leveraging LLMs for Scoring ENEM Essays. In STIL (Fortaleza/CE). SBC, Porto Alegre, RS, Brasil, 467–477. DOI: 10.5753/stil.2025.37847
Matheus Yasuo Ribeiro Utino, Elton H. Matsushima, Aline Paes, and Paulo Mann. 2025. Synthetic Data for Mental Health: A Comparative Analysis of LLMs, BERT, and Copy-Based Augmentation. In BRACIS. Springer Nature Switzerland, Cham, 393–408.
MengtingWan, Tara Safavi, Sujay Kumar Jauhar, Yujin Kim, Scott Counts, Jennifer Neville, Siddharth Suri, Chirag Shah, RyenW.White, Longqi Yang, Reid Andersen, Georg Buscher, Dhruv Joshi, and Nagu Rangan. 2024. TnT-LLM: Text Mining at Scale with Large Language Models. In KDD (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, New York, NY, USA, 5836–5847. DOI: 10.1145/3637528.3671647
Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In ICLR.
Yunyi Zhang, Minhao Jiang, Yu Meng, Yu Zhang, and Jiawei Han. 2023. PIEClass: Weakly-Supervised Text Classification with Prompting and Noise-Robust Iterative Ensemble Training. In EMNLP. Association for Computational Linguistics, Singapore, 12655–12670. DOI: 10.18653/v1/2023.emnlp-main.780
Yuwei Zhang, ZihanWang, and Jingbo Shang. 2023. ClusterLLM: Large Language Models as a Guide for Text Clustering. In EMNLP. ACL, Singapore, 13903–13920. DOI: 10.18653/v1/2023.emnlp-main.858
