Enhancing Table Retrieval with Dual Graph Representations

Tianyun Liu, Xinghua Zhang, Zhenyu Zhang, Yubin Wang, Quangang Li, Shuai Zhang, Tingwen Liu

Abstract: Table retrieval aims to rank candidate tables for answering natural language query, in which the most critical problem is how to learn informative representations for structured tables. Most previous methods roughly flatten the table and send it into a sequence encoder, ignoring the structure information of tables and the semantic interaction between table cells and contexts. In this paper, we propose a dual graph based method to perceive the semantics and structure of tables, so as to preferably support the downstream table retrieval task. Inspired by human cognition, we first decouple a table into the row view and column view, then build dual graphs from these two views with the consideration of table contexts. Afterward, intra-graph and inter-graph interactions are iteratively performed for aggregating and exchanging local row- and column-oriented features respectively, and an adaptive fusion strategy is eventually tailor-made for sophisticated table representations. In this way, the table structure and semantic information are well considered with dual-graph modeling. Consequently, the input query can match the target tables based on their full-fledged table representations and achieve the ultimate ranking results more accurately. Extensive experiments verify the superiority of our dual graphs over strong baselines on two table retrieval datasets WikiTables and WebQueryTable. Further analyses also confirm the adaptability for row-/column-oriented tables, and show the rationality and generalization of dual graphs. The source code is available at https://github.com/ty33123/DualG .