Data-to-text Generation with Data Control and Multi-loss Fusion

The goal of data-to-text generation is to generate fluent and fidelity natural language based on input data. At present, most models focus on data planning. Generating text according to the planning can effectively improve the fidelity to the generated text, but they lack attention to the fluency of the generated text, resulting in poor readability of the generated text. In order to address this issue, we propose a data-to-text generation model based on a pre-training language model. A control module is cleverly designed inside the language model to improve the fidelity to generated text. In training stage, multiple loss functions are cleverly designed to optimize the fluency and fidelity of generated text. Finally, the experimental results of RotoWire datasets show that the strong prior knowledge based on the Generative Pre-Training 2.0 (GPT2) can effectively improve the fluency of the generated text, and the proposed model in this paper can generate and control the fluency and fidelity of the generated text simultaneously.