FLAP: Table-to-Text Generation with Feature Indication and Numerical Reasoning Pretraining

Anonymous

Abstract: Recent neural models have shown success in table-to-text generation. However, the performance of content selection and content planning is still unsatisfactory. In this paper, we propose an effective framework with Feature indication and numericaL reAsoning Pretraining (FLAP) to help the neural generation model on content selection and planning. First, rather than treating the table as a sequence of token embeddings, we map each table into a numerical vector to utilize the real number information. We further propose a feature indication mechanism that introduces combination invariant bias to reduce the exposure bias problem in our generation system. Second, we propose a numerical reasoning pretraining task to help model do numerical reasoning upon the selected subset of tables. Experiments show that our framework outperforms the strong baselines on metrics of both content selection and planning on ROTOWIRE and RW-FG.