When More Metadata Misleads: Probing Field-Level Grounding in LLM-Generated Museum Descriptions

TL;DR

This paper challenges the assumption that more metadata improves grounding in LLM-generated museum descriptions. The study finds that metadata composition, including field type and image availability, significantly impacts generation behavior. The results highlight the need for field- and claim-level evaluation in metadata-conditioned data-to-text tasks.

Why It Matters

The findings have practical implications for the development of more accurate and reliable museum description generation systems. Improved grounding can enhance the overall quality and usefulness of generated descriptions, benefiting both museum curators and visitors. This research contributes to the advancement of data-to-text technologies in real-world applications.

Key Findings

• metadata quantity does not solely determine generation behavior
• field type influences grounding, omission, or inference of information
• images can support visual grounding but also increase unsupported interpretation