TableSF: A Structural Bias Framework for Table-To-Text Generation

Di Liu, Weihua Wang, Feilong Bao, Guanglai Gao

Abstract: Table-to-text generation is to generate a description from the tabular data. Existing methods typically encoded table content in a fixed order and relied heavily on the table row or column sequence. They generated error text descriptions when the row or column sequence changed. To solve the above problems, we proposed a novel structural bias framework that encodes tables using a modified self-attention mechanism. The framework captures the connectivity of cells in the same row or column through structural bias attention, distinguishing important cells from unimportant cells from a structural perspective. The structural bias attention will be added on top of the full self-attention, which can obtain the full structural information of the table. Experimental results show that this method generates better text descriptions on the public dataset and accomplishes a better understanding of the structured tables. This method not only obtains the relationship between cells but also improves the robustness of the pre-trained model.