Pragmatics Meets Culture: Culturally-adapted Artwork Description Generation and Evaluation
A new pragmatic speaker model improves AI's ability to explain art across cultures by 8.2%.
A team of researchers from the University of Maryland and Cornell Tech has published a novel paper, 'Pragmatics Meets Culture: Culturally-adapted Artwork Description Generation and Evaluation,' addressing a critical gap in AI. While language models are known for cultural bias in decision-making, this work explores their 'cultural familiarity' in open-ended generation. The researchers introduced a new task: generating artwork descriptions tailored for audiences from different cultural backgrounds who may not share the same symbolic or narrative context embedded in the art.
To tackle this, the team proposed an evaluation framework based on culturally-grounded question answering to measure a model's 'cultural competence.' Their key finding was that standard base models performed poorly. However, by implementing a 'pragmatic speaker model'—an AI that considers the listener's knowledge—they achieved an 8.2% improvement in simulated listener comprehension. This result was validated by a human study, where participants rated the pragmatically-enhanced model as 8.0% more helpful for understanding the artwork, confirming the real-world utility of their approach.
- Researchers introduced a new task and framework for evaluating AI's ability to generate culturally-adapted art descriptions.
- Base language models showed limited cultural competence, but a pragmatic speaker model boosted simulated comprehension by 8.2%.
- Human evaluators rated the pragmatic model as 8.0% more helpful, proving the method's effectiveness for real-world applications.
Why It Matters
This research provides a concrete method to reduce cultural bias in AI, making generative models more useful and accessible for global, cross-cultural communication.