Schema-Guided Natural Language Generation

  • 2020-05-11 23:01:22
  • Yuheng Du, Shereen Oraby, Vittorio Perera, Minmin Shen, Anjali Narayan-Chen, Tagyoung Chung, Anu Venkatesh, Dilek Hakkani-Tur
  • 1

Abstract

Neural network based approaches to natural language generation (NLG) havegained popularity in recent years. The goal of the task is to generate anatural language string to realize an input meaning representation, hence largedatasets of paired utterances and their meaning representations are used fortraining the network. However, dataset creation for language generation is anarduous task, and popular datasets designed for training these generatorsmostly consist of simple meaning representations composed of slot and valuetokens to be realized. These simple meaning representations do not include anycontextual information that may be helpful for training an NLG system togeneralize, such as domain information and descriptions of slots and values. Inthis paper, we present the novel task of Schema-Guided Natural LanguageGeneration, in which we repurpose an existing dataset for another task: dialogstate tracking. Dialog state tracking data includes a large and rich schemaspanning multiple different attributes, including information about the domain,user intent, and slot descriptions. We train different state-of-the-art modelsfor neural natural language generation on this data and show that inclusion ofthe rich schema allows our models to produce higher quality outputs both interms of semantics and diversity. We also conduct experiments comparing modelperformance on seen versus unseen domains. Finally, we present human evaluationresults and analysis demonstrating high ratings for overall output quality.

 

Quick Read (beta)

loading the full paper ...