Search Results

This paper investigates the ability of LLMs to identify culturally sensitive communication in European Portuguese, in comparison with human performance. The study addresses cross-cultural communication issues arising within a single language, including miscommunication, bias, framing, and stereotyping. The task is formulated as a multi-component problem combining detection, classification, span identification, and reformulation of problematic content. A dataset of 60 sentences, balanced between problematic and neutral instances, was constructed from real-world corpora and annotated by expert linguists. Three open-weight multilingual models were evaluated alongside human annotators with diverse backgrounds. Results show that LLMs achieve high accuracy in detecting problematic content, outperforming non-expert participants and approaching expert performance. However, both models and humans exhibit low agreement in the classification of communication issues, reflecting the conceptual overlap between categories. While models generate consistent reformulations, they show limitations in span precision and contextual interpretation. The findings highlight a gap between detection and interpretation capabilities and support the use of LLMs as assistive tools in cross-cultural communication tasks, particularly in politically sensitive contexts.