A Two-Stage Table-to-Text Generation Method Based on Structure Understanding and Language Generation

Despite significant advancements in structured data to text generation by neural networks, the heterogeneity of datasets has limited the research in this field. Large pre-trained models have shown great potential in addressing this issue. However, the challenge of effectively leveraging table structure information to provide richer content for pre-trained models and bridging the gap between structured tables and textual input remains underexplored. Current research predominantly focuses on processing unstructured text, such as natural language sentences, documents, and news articles, while structured data, such as tables, charts, and databases, have not been sufficiently studied. To fill this gap, we formulate a two-stage approach for table structure understanding and text generation. Specifically, we designed a novel self-supervised learning objective to enhance the model's awareness of table structures, thereby creating a structure-aware text generation model that overcomes the limitations of existing models in handling linearized input. Furthermore, a multi-channel decoder framework was adopted to enhance the fluency and fidelity of table-descriptive text generation. Experimental assessments on two benchmark datasets, coupled with human evaluations, validate that our approach can generate both faithful and coherent descriptive text for various table structures.