Multi-Domain Task-Completion Dialog Challenge

As part of the Eighth Dialog System Technology Challenge (DSTC8), Microsoft Research and Tsinghua University are hosting a track intended to foster progress in two important aspects of dialog systems: dialog complexity and scaling to new domains. For this DSTC8 track, there are two tasks you can compete in (see below). The challenge runs from June 17, 2019 – October 6, 2019.

Task 1 –  There is increasing interest in building complex bots that span over multiple sub-domains to accomplish a complex user goal such as travel planning. Travel planning may include sub-domains like hotels, restaurants, tourist attractions, and so on. To advance state-of-the-art technologies for handling complex dialogs, we offer a timely task focusing on multi-domain end-to-end task completion dialog.

Sign up to participate in Task 1 at https://aka.ms/dstc8-task1.

Task 2 – Neural dialog systems require very large datasets to learn how to output consistent and grammatically-correct sentences. This need for large datasets makes it extremely hard to scale out the system to new domains with limited in-domain data. With Task 2, our goal is to investigate whether sample complexity can decrease with time. In other words, the goal of Task 2 is to investigate whether a dialog system that was trained on a large corpus can learn to converse about a new domain given a much smaller in-domain corpus.

Sign up to participate in Task 2 at https://aka.ms/dstc8-task2.

Check out the MetaLWOz dataset that is used for Task 2.