An LLM-based methodology for the automatic detection of bias in the DuoWikiBias corpus.

Bias detection remains a challenge in Natural Language Processing, particularly in non-English contexts, due to the conceptual ambiguity of bias and the scarcity of annotated resources. This study addresses the lack of Spanish-language resources by investigating the automatic detection of framing, epistemological, and demographic biases. We introduce DuoWikiBias, a novel parallel corpus derived from Wikipedia for Spanish bias classification. We evaluate Large Language Models (Llama and Gemma) using advanced prompting techniques-CARP and Metacognition-combined with a Gradient Ascent unlearning method to refine model attention. Their performance is compared against classical approaches, including logistic regression with S-BERT embeddings and linguistic features. Results show that advanced prompting substantially improves performance over simple instructions, while the best overall performance (F1 = 0.796) is achieved by combining CARP-based features with Gradient Ascent and a Support Vector Machine classifier. These findings suggest that LLMs are effective for bias-aware representation learning, but hybrid approaches with traditional classifiers remain competitive. This work provides both a validated dataset and a methodological framework for bias detection in Spanish NLP.