Reasoning Knowledge Transfer for Logical Table-to-Text Generation
Baoqiang Liu, Yu Bai, Fang Wei Cai, Shuang Xue, Na Ye, Xinyuan Ye · 2024
Logical table-to-text generation (LT2T) aims to generate logically faithful textual descriptions from tables. However, existing end-to-end LT2T models, which directly utilize descriptions as learning objectives, often struggle to ensure logical faithfulness due to the absence of a formal reasoning process. To solve this problem, we introduce reasoning knowledge transfer, a framework designed to transfer the reasoning knowledge from external dataset and integrate the reasoning knowledge into the generation of descriptions. Our framework fine-tunes a transfer model on external dataset to transfer reasoning knowledge and trains a knowledge-driven generation model by using transferred reasoning knowledge with two self-supervised objectives: logical reasoning and logical summary. Our framework can align the table description with the reasoning knowledge and generate more logically faithful descriptions. Experimental results show the effectiveness of our method, demonstrating significant improvements of 1.9 in SP-Acc and 1.2 in NLI-Acc over the current state-of-the-art model.