Autor(es): Gómez, Leo Ebert
Data: 2025
Identificador Persistente: http://hdl.handle.net/10362/202067
Origem: Repositório Institucional da UNL
Assunto(s): Large language models; Hierarchical classification; E-Commerce; In-context learning; Fine tuning; Prompt engineering; Knowledge graphs; Retrieval augmented generation; Entity matching
This research explored techniques to improve Large Language Models performance for Hierarchical Product Classification (HPC), including optimized fine-tuning, optimal prompting techniques, taxonomy-specific Knowledge Graphs, leveraging Retrieval-Augmented Generation, and implementing LLM-based Entity Matching. Tested on benchmark datasets Icecat and WDC-222, these methods significantly enhanced LLMs’ ability to solve HPC tasks across various scenarios. Results achieved a hierarchical F1-score (hF) of 0.921, surpassing traditional DL benchmarks (0.85 hF). While not outperforming proprietary models like GPT, the proposed approaches offer a cost-efficient and effective alternative for businesses, demonstrating strong performance without reliance on expensive LLM solutions.
