Data-to-text Generation with Macro Planning

Abstract Recent approaches to data-to-text generation have adopted the very successful encoder-decoder architecture or variants thereof. These models generate text that is fluent (but often imprecise) and perform quite poorly at selecting appropriate content ordering it coherently. To overcome some of these issues, we propose a neural model with macro planning stage followed by reminiscent traditional methods which embrace separate modules for surface realization. Macro plans represent high level organization important such as entities, events, their interactions; they are learned from data given input generator. Extensive experiments on two benchmarks (RotoWire MLB) show our approach outperforms competitive baselines in terms automatic human evaluation.