Studio Ousia’s AI Defeats a Team of Human Quiz Champions


TOKYO - December 14, 2017 - Studio Ousia Inc. is excited to announce that on December 8, its artificial intelligence defeated a team of human quiz champions at the “Human Computer Question Answering Competition” held at NIPS 2017 (Conference on Neural Information Processing Systems, Long Beach, California - December 4-9, 2017). NIPS is a top academic conference for machine learning. Studio Ousia scored more than double the combined human team score (465:200).

(Interim scores: score for the human team on the left and for the AI on the right)

The competition was held in a quiz bowl format. One of the typical characteristics of a quiz bowl is the lockout buzzer system. This means the quicker the players respond (i.e. with less information), the more points they receive. While it is challenging for humans, it is even more so for computers. The organizers of this “Human Computer Question Answering Competition” are top researchers in the field, coming from Maryland University, Stanford University, University of Colorado, and the Allen Institute for AI.

AI systems from various institutions first competed to determine which machine was the champion. Studio Ousia’s AI won the championship and, therefore, proceeded to compete against a team of human quiz champions. The human team consisted of six top quiz experts including Raj Dhuwalia, a Jeopardy! champion and winner of 250k on “Who Wants to be a Millionaire”; David Farris, a mathematician and three-time national champion, and Charles Meigs, the two-time D2 quiz bowl All-Star.

While it is common knowledge that in 2011, IBM’s Watson defeated Ken Jennings, a Jeopardy! champion, Studio Ousia’s recent AI performance is so remarkable because our machine was able to beat six top human quiz experts playing together.

Studio Ousia managed to achieve this result by applying two deep learning-based NLU (Natural Language Understanding) models that were trained using past quiz bowl questions as well as encyclopedic information from Wikipedia. The company also used ensemble learning algorithms (gradient boosted regression trees) to combine signals obtained from NLU and other conventional models. See slides for further details:

The technology employed during the competition was developed leveraging the knowledge acquired from R&D performed by Studio Ousia for their automated answering product, QA ENGINE; a product used primarily for helpdesk automation. Studio Ousia plans to further improve the performance and usability of QA ENGINE, aiming to become the de facto standard for question answering systems around the world.



Studio Ousia Company Information:

  • A deep learning company focused on natural language processing
  • Products: QA Engine (question answering system for helpdesk automation), Semantic Kernel (entity linking API)
  • Founders: Yasuhiro Watanabe (CEO), Ikuya Yamada (CTO)
  • Date of foundation: Feb 7, 2007
  • Capital: ¥187 Mil.
  • Share holders: Founders and employees, NID, Samsung Venture Investment, STCP
  • HQ Address: 3-27-15, Shibuya-ku, Jingu-mae, Tokyo Japan 150-0001
  • Company web site:
  • QA ENGINE web site:

Contacts: info (at)