Rethinking Table Retrieval from Data Lakes

Table retrieval from data lakes has recently become important for many downstream tasks, including data discovery and table question answering. Existing table retrieval approaches estimate each table's relevance to a particular information need and return a ranking of the most relevant tables. This approach is not ideal since (1) the returned tables often include irrelevant data and (2) the required information may be scattered across multiple tables. To address these issues, we propose the idea of fine-grained structured table retrieval and present our vision of R2D2, a system which slices tables into small tiles that are later composed into a structured result that is tailored to the user-provided information need. An initial evaluation of our approach demonstrates how our idea can improve table retrieval and relevant downstream tasks such as table question answering.