Complex Table Question Answering with Multiple Cells Recall Based on Extended Cell Semantic Matching

Abstract

Tables, as a form of structured or semi-structured data, are widely found in documents, reports, and data manuals. Table-based question answering (TableQA) plays a key role in table document analysis and understanding. Existing approaches to TableQA can be broadly categorized into content-matching methods and end-to-end generation methods based on encoder–decoder deep neural networks. Content-matching methods return one or more table cells as answers, thereby preserving the original data and making them more suitable for downstream tasks. End-to-end methods, especially those leveraging large language models (LLMs), have achieved strong performance on various benchmarks. However, the variability in LLM-generated expressions and their heavy reliance on prompt engineering limit their applicability where answer fidelity to the source table is critical. In this work, we propose CBCM (Cell-by-Cell semantic Matching), a fine-grained cell-level matching method that extends the traditional row- and column-matching paradigm to improve accuracy and applicability in TableQA. Furthermore, based on the public IM-TQA dataset, we construct a new benchmark, IM-TQA-X, specifically designed for the multi-row and multi-column cell recall task, a scenario underexplored in existing state-of-the-art content-matching methods. Experimental results show that CBCM improves overall accuracy by 2.5% over the latest row- and column-matching method RGCNRCI (Relational Graph Convolutional Networks based Row and Column Intersection), and boosts accuracy in the multi-row and multi-column recall task from 4.3% to 34%.