Multi-table Question Answering Method Based on Correlation Evaluation and Precomputed Cube

The goal of multi-table question answering (Multi-TableQA) is to enable intelligent extraction and understanding of structured table data via natural language queries. The most commonly used autoregressive pretraining language models (PLMs) encounter problems such as noise sequence information, sequence length and numerical computation limitation. To address these issues, we propose two preprocessing steps: a correlation evaluator and a precomputed cube (TaCube). The correlation evaluator assesses the relevance between questions and table columns, discarding irrelevant columns to minimize information noise and reduce sequence length. TaCube precomputes aggregation and arithmetic operations based on the query, thereby equipping the model with rapid access to numerical computations and enhancing the PLM’s numerical reasoning capabilities. Experimental results demonstrate that these preprocessing steps increase the exact match accuracy on the Spider and GeoQuery multi-table question answering datasets to 33.99% and 61.43%, respectively, outperforming existing methods.