Structure-aware Learning for Table-to-text Generation

The task of table-to-text generation is to generate the text describing the semantic facts defined by key-value pairs in a table. The existing work based on pre-trained models are challenged by the semantic relationships between keys and values in tables. In this paper, we propose a structure-aware learning-based table-to-text generation model. The method includes (1) a definable mask matrix and an automatic generation algorithm for the definition of semantic relations of table keys and values; (2) a structure-aware multi-headed self-attention mechanism introducing the mask matrix for the learning of table structural relations; (3) a training method for the text reconstruction task to enhance the learning of semantic structural representation of tables. Experimental results on the E2E and WikiBio datasets show that the proposed method improves 1.17 and 0.63 BLEU values, respectively, compared with the strong baseline model T5.