Neural Data-to-Text Generation Guided by Predicted Plan

Abstract

Data-to-text generation task aims to generate natural language text from structured data and has made great progress in recent years with the help of end-to-end neural network models. However, these end-to-end approaches often ignore the structure of the output text and convey the information in the input data in a random order. When faced with the data-to-text generation task, a person tends to make a plan for the complex input before writing the final text, which is inconsistent with end-to-end approaches. In this paper, we propose a novel plan-guided data-to-text generation framework consisting of a plan generator GT5 and a text generator Share-T5. The plan generator GT5 first predicts a plan based on the input data and then the text generator Share-T5 generates the target text based on the input data and the predicted plan. Empirical comparisons with strong baselines on two benchmark datasets show that our proposed plan-guided data-to-text generation framework can significantly improve the performance of plan prediction and text generation.