📝 SummarXiv

Abstract

Open data portals are essential for providing public access to open datasets. However, their search interfaces typically rely on keyword-based mechanisms and a narrow set of metadata fields. This design makes it difficult for users to find datasets using natural language queries. The problem is worsened by metadata that is often incomplete or inconsistent, especially when users lack familiarity with domain-specific terminology. In this paper, we examine how individual metadata fields affect the success of conversational dataset retrieval and whether LLMs can help bridge the gap between natural queries and structured metadata. We conduct a controlled ablation study using simulated natural language queries over real-world datasets to evaluate retrieval performance under various metadata configurations. We also compare existing content of the metadata field 'description' with LLM-generated content, exploring how different prompting strategies influence quality and impact on search outcomes. Our findings suggest that dataset descriptions play a central role in aligning with user intent, and that LLM-generated descriptions can support effective retrieval. These results highlight both the limitations of current metadata practices and the potential of generative models to improve dataset discoverability in open data portals.