arXiv (Cornell University) · 2019 · 43 citations · 0 references
ComputingComputer ScienceComputational LinguisticsMachine LearningNatural Language ProcessingLanguage ModelingNatural Language GenerationText GenerationText RepresentationAutomated Text GenerationData-to-text GenerationText Structure Knowledge
Data-to-text generation can be conceptually divided into two parts: ordering\nand structuring the information (planning), and generating fluent language\ndescribing the information (realization). Modern neural generation systems\nconflate these two steps into a single end-to-end differentiable system. We\npropose to split the generation process into a symbolic text-planning stage\nthat is faithful to the input, followed by a neural generation stage that\nfocuses only on realization. For training a plan-to-text generator, we present\na method for matching reference texts to their corresponding text plans. For\ninference time, we describe a method for selecting high-quality text plans for\nnew inputs. We implement and evaluate our approach on the WebNLG benchmark. Our\nresults demonstrate that decoupling text planning from neural realization\nindeed improves the system's reliability and adequacy while maintaining fluent\noutput. We observe improvements both in BLEU scores and in manual evaluations.\nAnother benefit of our approach is the ability to output diverse realizations\nof the same input, paving the way to explicit control over the generated text\nstructure.\n