Recipes for Building an Open-Domain Chatbot

Royal Sequeira · May 15, 2021

When was the last time you had a great conversation? A good conversation requires various skills: engaging talking points, exhibiting empathy etc. This article explores the the development of large scale models that can learn these skills for a open-domain chat. For more details, read the original paper.

Task

A great conversational chatbot need to have smooth conversations that include the following traits: following the interlocuter and providing new points for discussions, exhibiting prior knowldege, empathy, personality, all while maintaining a consister persona.

Metrics

  • Automatic Evaluations: The automatic evaluation of the individual components included 4 different metrics:
    • Retriever: hits@1 for depth K (hits@1/K). This implies recall@1 while ranking among K candidates
    • Generator:
    • Retrieve and Refine (RetNRef):
    • Safety:
  • Self-chat Evaluations:

  • Full Evaluations:
\[\begin{equation} H(x) = -\sum_i p(x_i)\:log(p(x_i))\end{equation}\]

Twitter, Facebook