Abstract

Natural language generation lies at the core of generative dialogue systemsand conversational agents. We describe an ensemble neural language generator,and present several novel methods for data representation and augmentation thatyield improved results in our model. We test the model on three datasets in therestaurant, TV and laptop domains, and report both objective and subjectiveevaluations of our best model. Using a range of automatic metrics, as well ashuman evaluators, we show that our approach achieves better results thanstate-of-the-art models on the same datasets.

 

Quick Read (beta)