IBM's debating system, Project Debater, is the first AI system to engage in complex debates with humans.
Human debater Harish Hatarajan is one of the best debaters. He has won the most debating contests, including the 2012 European debating championship. Moreover, the human debater has never competed with IBM's machine debater.
The debate was held at a public conference centre in San Francisco with 800 people and broadcast live around the world.
Whether it's an AI system or an opponent, there's no preparation for the debate in advance, so who can convince you depends entirely on the debate.
Debate: Should Preschool Education be Supported?
SquareShould support (machines)
con sideNo support (human debater)
formFifteen minutes before the formal debate, the debate will be published. The machine will first state for three minutes and the human will state for three minutes, totaling three rounds.
ResultSeventeen percent of people believe that what human beings say is justified after the end of the campaign, turning from supporting the right side to supporting the opposite side. Nearly 60% of people think that in the process of debating machines, they have gained knowledge.
Two thousand years ago, debate and rhetoric had attracted the greatest Greek philosophers, and the ability to debate and express reasoning to others laid the foundation of human civilization. This is a decisive feature in the human sense. Therefore, the development of AI systems in this field, which may be unique to human beings, is a formidable challenge.
At the beginning of each debate, debaters choose a controversial topic, and IBM machine debaters create a coherent four-minute speech by choosing the most relevant and persuasive arguments and evidence from the 10 billion sentences in authoritative newspapers and journals.
Subsequently, the machine debater listens to the opponent's four-minute response, identifies the key propositions, and immediately generates convincing refutations, demonstrating human-like reasoning. Before or after the debate, the audience is asked to vote. After the debate, both sides make brief statements.
Machine debaters represent a new AI challenge because debates are more open than playing chess or Go.
Even in competitive debates, some rules originate from the ancient culture of human discussion, unlike the rules of chess with clearly defined actions. This fundamental difference has important implications. First, in complex board games, AI systems may come up with any strategy to ensure victory, even if the relevant actions cannot be easily understood by humans. This is no longer the case in the debate. Artificial intelligence system must adapt to human rationality, and put forward the reasoning route that human can follow and sympathize with.
Second, in sharp contrast to previous game-related challenges, AI systems have no natural scoring function in debates. The value of individual action, that is, argument, is often inherently subjective; moreover, there is not even an agreed objective measure to determine who is the winner of the debate.
Machine debaters have proved that AI can play an important role in this unknown field. We believe that it will make a new form of decision-making possible. It will synergistically combine human and machine to make decisions under more abundant information.
How would you quantify the value of an opinion? Or the value of some supporting evidence? One of the challenges facing researchers is to work in an area where there are no correct or wrong answers, and there is very little markup data available.
In order to develop machine debaters, IBM research teams must give the system three functions, each of which is a new breakthrough in artificial intelligence.
First, data-driven speech writing and presentation: Machine debater is the first computer case that can digest a large corpus. It can briefly describe a controversial topic, write well-structured speeches, deliver speeches in a clear and purposeful way, and incorporate humor in appropriate situations.
Second, listening and understanding: the ability to identify key concepts and perspectives hidden in long-term continuous oral English.
Thirdly, simulate the logic of human debate: simulate the world of human controversy and controversy in a unique way of knowledge representation, so that AI system can put forward principled arguments according to needs.
Last June, IBM demonstrated the machine debater in a controlled environment. Two human and machine debaters debated, with a total audience of 50. These two human debaters have previously competed with machine debaters.