Prompting or Fine-tuning? Evaluating Relation Classification in Portuguese for Knowledge Graph Construction

This paper explores relation classification as a step toward building structured knowledge graphs from unstructured Portuguese text. We evaluate prompt-based approaches using generative large language models and compare them with a fine-tuned BERT model. Experiments are conducted on a custom relation extraction dataset built by aligning Wikidata triples with sentences from Wikipedia in Portuguese, a lower-resourced language than English. Results show that while prompt-based methods offer flexibility and reasonable performance, fine-tuning yields significantly better results in identifying relation types, an essential aspect for accurate and reliable knowledge graph construction.