Grounded response generation task at dstc7

This paper offers and overview of the “Sentence Generation” task (Task 2) at the 7th Dialog System Technology Challenges (DSTC7). In this task, the goal is to generate conversational responses that go beyond chitchat, by producing informational responses that are grounded in external knowledge following the framework proposed by Ghazvininejad et al.[1]. We argue that such a grounding can empower end-to-end dialogue modeling research, so far overwhelmingly devoted to chit-chat, to produce informative and ultimately more useful dialogues. We received 26 system submissions from 7 teams, and evaluated them using both automatic and human evaluation. We discuss the outcomes of these two evaluations, noting the merits and drawback of each automatic metric based on their correlation with human judgments. Finally, we briefly outline challenges for future work in grounded response generation.