This paper addresses the problem of table retrieval in open-domain question-answering systems, where retrieving multiple tables and joining them through a join plan is often necessary. Previous methods focus on retrieving a single table or multiple tables identified through question decomposition, but these approaches are insufficient as many questions require joining tables that cannot be discerned from the user query. To address this, the authors propose a method that uncovers useful join relations during table retrieval. They use a novel re-ranking method formulated as a mixed-integer program that considers both table-query relevance and table-table relevance. The method outperforms state-of-the-art approaches in table retrieval and end-to-end QA by up to 9.3% and 5.4% in F1 score and accuracy, respectively. The paper introduces a re-ranking approach that considers both query-table relevance and table-table relevance to select the best set of tables during test time. The method is evaluated on Spider and Bird datasets, demonstrating improved retrieval and end-to-end performance. The paper also discusses the importance of considering table relationships during retrieval and the challenges of inferring join relationships in real-world scenarios. The authors conclude that both fine-grained query-table relevance and table-table relevance are essential for better retrieval performance. The paper also highlights the effectiveness of their approach in handling complex and realistic queries involving multiple tables.This paper addresses the problem of table retrieval in open-domain question-answering systems, where retrieving multiple tables and joining them through a join plan is often necessary. Previous methods focus on retrieving a single table or multiple tables identified through question decomposition, but these approaches are insufficient as many questions require joining tables that cannot be discerned from the user query. To address this, the authors propose a method that uncovers useful join relations during table retrieval. They use a novel re-ranking method formulated as a mixed-integer program that considers both table-query relevance and table-table relevance. The method outperforms state-of-the-art approaches in table retrieval and end-to-end QA by up to 9.3% and 5.4% in F1 score and accuracy, respectively. The paper introduces a re-ranking approach that considers both query-table relevance and table-table relevance to select the best set of tables during test time. The method is evaluated on Spider and Bird datasets, demonstrating improved retrieval and end-to-end performance. The paper also discusses the importance of considering table relationships during retrieval and the challenges of inferring join relationships in real-world scenarios. The authors conclude that both fine-grained query-table relevance and table-table relevance are essential for better retrieval performance. The paper also highlights the effectiveness of their approach in handling complex and realistic queries involving multiple tables.