AI

Watch IBM’s AI System Debate a Human Champion Live at Think 2019

Share this post:

Great public debates have sparked our imagination since the days of ancient Greece. This intellectual tradition will take on new life at the IBM Think conference in San Francisco, when IBM Research and Intelligence Squared U.S. host a live public debate on Monday, February 11 between a human and an AI.

At the center is IBM Project Debater, the first AI system that can debate humans on complex topics. To do this effectively, the system must gather relevant facts and opinions, form them into structured arguments, and then use precise language in a clear and persuasive way.

Project Debater’s first live public debate took place in June before a small audience. At Think, Project Debater will face off against a champion debater in front of a large in-person audience, with many more watching via livestream (watch here on Feb. 11 at 5 pm PT).

Harish Natarajan, Project Debater’s opponent at Think 2019, is 2016 World Debating Championships Grand Finalist and 2012 European Debate Champion. Harish holds the world record for most debate competition victories.

Project Debater’s opponent is 2016 World Debating Championships Grand Finalist and 2012 European Debate Champion Harish Natarajan. Harish holds the world record for most debate competition victories. Each side has only 15 minutes to prep for the debate, during which they prepare arguments for and against a given thesis statement. Both sides will then present a four-minute opening statement, a four-minute rebuttal, and a two-minute summary. The exhibition promises to be lively, with attendees including top debaters from Bay Area schools and more than 100 journalists.

Harish is the highest caliber opponent Project Debater has faced, so we expect the technology will be pushed to its limits. Still, it may be that some in the audience don’t fully grasp the difficulty of preparing AI for such a challenge. Here’s more on what goes into a debate:

  • Project Debater’s knowledge base consists of around 10 billion sentences, taken from newspapers and journals.
  • In a live debate, Project Debater debates a topic that it has never trained on given a very short sentence describing the motion. The first step is to build an opening speech to defend or oppose this motion. Project Debater searches for short pieces of text in the massive corpora that can serve this purpose. This requires a deep understanding of human language and its infinite nuances and very precise stance identification, something that is not always easy for humans and is certainly very difficult for computers.
  • This process can result in a few hundred relevant text segments. In order to debate effectively, the system needs to construct the strongest and most diverse arguments to support its case. Project Debater does this by removing redundant argumentative texts, selecting the strongest remaining claims and evidence, and arranging these by theme, creating the base of the narrative to support or contest the motion.
  • It also uses a knowledge graph that allows it to find arguments to support the general human dilemmas that are raised by the debate topic, such as when it is right for the government to coerce its citizens, infringing on their personal freedom of choice.
  • Project Debater pieces all the selected arguments together to create a persuasive speech that lasts approximately four minutes. This process only takes a few minutes. Then, it’s ready to deliver its opening speech.
  • The next step is to listen to the opponent’s response, digest it and build the rebuttal. Generating a good rebuttal is the most challenging part of debate for both humans and machines. Project Debater applies many techniques, including those to anticipate and identify the opponent’s arguments. It then aims to respond with claims and evidence that counter these arguments.

These capabilities have been developed over the course of the project and our team has published more than 30 research papers to document and share what we’ve learned.

Though the format and challenge of debate have allowed us to shape Project Debater’s capabilities, we envision a future for the technology beyond the podium. It could be used, for instance, to promote more civil debates in online comment forums or by an attorney preparing for a trial where it could review legal precedents and test the strengths and weaknesses of a case using a mock legal debate. In the financial services industry, Debater could identify financial facts that either support or undermine an investment strategy. Or it could be applied as a voice interaction layer for various complex customer experiences, or even to improve critical thinking and critical writing skills of youngsters.

Debate is all about language. Mastering human language is one of the most ambitious goals of AI. Project Debater moves us one step closer in this journey. In the grand scheme, Project Debater reflects the mission of IBM Research to develop broad AI that learns across different disciplines to augment human intelligence. It absorbs large sets of information and perspectives in the quest for one simple goal: to help us make better and more informed decisions.

If you’re interested in experiencing the AI technology yourself, check out Project Debater – Speech by Crowd, which uses the core AI behind Project Debater to collect arguments from large audiences and automatically construct persuasive viewpoints to support or contest a topic such as, “we should ban the sale of violent video games to minors.” After Monday’s debate, we’ll post a new topic and hope you’ll make your voice heard.

Click here to watch Project Debater live on Monday, Feb. 11 from 5-6 pm PST. During this hour you will have the opportunity to watch the full live debate and hear from the scientists behind the technology. John Donvan, four-time Emmy Award-winner and host of the Intelligence Squared U.S. debate series, will moderate the session.

IBM Research

Noam Slonim

Distinguished Engineer, Project Debater, IBM Research-Haifa

More AI stories

Learning and Characterizing Causal Graphs with Latent Variables

Researchers from the MIT-IBM Watson AI Lab developed a new approach to characterize the set of plausible causal graphs from observational and interventional data that has latent variables.

Continue reading

More Is Less: Learning Efficient Video Representations

IBM researchers developed a novel low memory footprint and efficient architecture for spatio-temporal analysis of video. The results show strong performance on several benchmarks – and allow training of deeper models using larger sequences of input frames, which will lead to higher accuracy on video action recognition tasks.

Continue reading

Sobolev Independence Criterion

At NeurIPS 2019, IBM Research AI will showcase the “Sobolev Independence Criterion," a novel interpretable dependency measure that provides feature importance scores that can be used for controlling False Discovery Rate.

Continue reading