Retrieve Only Relevant Tables Whether Few or Many: Adaptive Table Retrieval Method

ACL ARR 2025 May Submission3624 Authors

Abstract:

Retrieving relevant tables from extensive databases for a given query is essential for accurately answering questions in tasks such as text-to-SQL and open-domain table question answering.
Top-$k$ retrieval strategies based on similarity scores between the query and tables are common in existing table retrieval methods.
However, the number of required tables varies across queries and cannot be known in advance.
Such strategies may either retrieve an undersized set of tables, preventing the model from gathering all that was needed, or retrieve too large a pool of tables, leading to the incorporation of unnecessary ones.
To address this issue, we present an adaptive table retrieval method that adjusts the number of tables retrieved according to the requirements of each query.
We adopt an adaptive thresholding mechanism to selectively retrieve tables and integrate a sliding-window re-ranking algorithm to efficiently process large candidate sets.
Extensive experiments on Spider, BIRD, and Spider 2.0 show that our method effectively addresses the limitations of the top-$k$ retrieval strategies, improving performance in both retrieval and downstream tasks.
Our code and data are available at https://anonymous.4open.science/r/Adaptive-Table-Retrieval.

Paper Type: Long

Research Area: Information Retrieval and Text Mining

Research Area Keywords: dense retrieval, re-ranking

Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 3624