IBM’s New AI System Debates Intellectuals Using ‘Argument Mining’

IBM's AI program debated Harish Natarajan during an IntelligenceSquared Forum. Photo courtesy of IntelligenceSquared.

IBM supercomputers are using “argument mining” to replicate human reason and argumentative frameworks.

In a new paper published by the scientific journal Nature, IBM’s team details the system architecture built for its ‘Project Debater’ AI program, which in 2019 went head-to-head with debate champion Harish Natarajan.

“Given the variety of tasks required to engage in a debate, it seems implausible to envision a monolith solution in the form of an end-to-end system, such as a neural model,” write the researchers. “Instead, our approach was to break the problem into modular tangible tasks pursued in parallel.”

Both Natarajan and Project Debater were given 15-minutes to prepare for a live debate on whether prekindergarten should be subsidized. Although the audience present at the IntelligenceSquared event found Natarajan won the debate, the AI program made rational arguments over the course of the discussion.

According to IBM, the four main modules comprising Project Debater’s architecture include argument mining, the Argument Knowledge Base (AKB), argument rebuttal, and debate construction.

The argument mining component draws offline from a database of over 400 million newspaper articles provided by Lexis Nexis and Wikipedia, breaking them down into sentences with key words which are further categorized.

“In the offline stage, once given a motion the system relies on this index to perform corpus-wide sentence-level argument mining, retrieving claims and evidence of related to the motion,” writes IBM, noting it retrieves and ranks certain arguments. “The argument mining module also searches for arguments that support the other side, aiming to prepare a set of claims the opponent may use and evidence that may serve as responses.”

The program’s AKB focuses on debate structure, and contributes to the rebuttal component by mapping principled arguments and counter-arguments to refute them.

“Claims coming from argument mining are rebutted with mined evidence texts that mention similar concepts and are predicted to oppose the opponent’s stance,” says IBM’s team.

IBM’s supercomputer Deep Blue famously beat former world chess champion Garry Kasparov in 1996.