Key points are not available for this paper at this time.
Data-to-text generation can be conceptually divided into two parts: ordering structuring the information (planning), and generating fluent language the information (realization). Modern neural generation systems these two steps into a single end-to-end differentiable system. We to split the generation process into a symbolic text-planning stage is faithful to the input, followed by a neural generation stage that only on realization. For training a plan-to-text generator, we present method for matching reference texts to their corresponding text plans. For time, we describe a method for selecting high-quality text plans for inputs. We implement and evaluate our approach on the WebNLG benchmark. Our demonstrate that decoupling text planning from neural realization improves the system's reliability and adequacy while maintaining fluent. We observe improvements both in BLEU scores and in manual evaluations. benefit of our approach is the ability to output diverse realizations the same input, paving the way to explicit control over the generated text.
Moryossef et al. (Sat,) studied this question.