End-to-End Recurrent Entity Network for Entity-Value Independent Goal-Oriented Dialog Learning

Published in Dialog System Technology Challenges 6, 2017

[PDF] [Code]


@article{wuend,
  title={End-to-End Recurrent Entity Network for Entity-Value Independent Goal-Oriented Dialog Learning},
  author={Wu, Chien-Sheng and Madotto, Andrea and Winata, Genta Indra and Fung, Pascale}
}

Abstract

This paper presents an end-to-end solution for the goal-oriented dialog system task in Dialog System Technology Challenges 6 (DSTC6). The challenge consists in learning a dialog policy from a given restaurant booking domain. End-to-end models are required to reason over dialog entities and to track the dialog states. Hence, we introduce a practical entity-value independent framework based on Recurrent Entity Networks. The framework is able to abstract linguistic entity by using a delexicalization mechanism, which improves the original model performance especially in test sets with out-of-vocabulary entities. Recurrent Entity Networks also plays an important role to represent the latent dialog state and the dialog policy. As shown in experiments, our framework can achieve a promising average Precision-1 of 96.56% in all the test sets.