IBM最新人工智能技术助力美国电视节目

Is it time to redistribute the wealth?

That's the topic two former Clinton administration officials, a former Greek finance minister and a researcher who has studied the economics of prostitution, explored in a televised debate—with a little help from IBM's latest artificial intelligence software.

The debate, which aired on Bloomberg's cable television channel in the U.S. New York time, is the debut episode of a series called "That's Debatable." Each episode will feature a debate on a different topic. The show is produced by the media and events company Intelligence Squared and Bloomberg, with sponsorship from Big Blue. John Donvan, an Emmy-winning journalist who has moderated debates for Intelligence Squared since 2008, is serving as the show's host and moderator.

The first debate featured four esteemed economists: Larry Summers, the former U.S. Treasury secretary and former president of Harvard University, Robert Reich, the former U.S. Labor secretary, Yanis Varoufakis, the outspoken former Greek finance minister, and Allison Shrager, a senior fellow at the Manhattan Institute known for her 2019 book, An Economist Walks Into a Brothel: And Other Unexpected Places to Understand Risk.

But IBM is no doubt hoping that, at least to some extent, its Watson A.I. software will have stolen the show. In particular, the television program highlights one of Watson's newest capabilities: categorizing and summarizing thousands, or even hundreds of thousands, of individual comments and opinions, and distilling them down to a handful of key points.

Called "key point analysis," the capability has grown out of IBM's "Project Debater" research, spearheaded by a team in its A.I. lab in Israel, which has involved building A.I. software capable of successfully debating humans—and helping humans surface strong arguments from a variety of different types of sources.

IBM hopes that it will be able to sell the technology to companies as a new way to conduct market research and solicit views from both employees and customers. The company also thinks the technology could be used by governments to better understand the views of citizens.

"Topic clustering and argument generation, those capabilities came from Debater and these are now being used with select customers," Dakshi Agrawal, IBM's chief architect for AI, said. He said key point analysis will allow businesses to "collect tens fo thousands of data points and distill them down to make more data-driven decisions."

Big Blue on the small screen

The company has staged a series of "grand challenge" public events in the past three years designed to showcase the A.I.-based technology, of which the "It's Debatable" T.V. program is the latest.

The key point summarization tool being featured in the television show can analyze thousands of comments that users submit through a website in response to a question, in this case, the debate proposition: "It is time to redistribute the wealth."

Using natural language processing—the kind of A.I. that can analyze and to some extent "understand" language—the IBM system scores these comments for relevancy, discarding those it sees as not being germane to the topic. Then it groups the remaining ones into two broad categories: those that support the proposition, and those that oppose it.

It then further groups the comments in each camp into a handful of key points, using its language processing algorithm to summarize the essence of each point and avoiding repetition of points that are simply expressing the same idea using different language.

The software also tells a user know how often each key point was mentioned by those submitting comments. "Actually conveying the prevalence of each keypoint in the data, this is important for decision-makers," Noam Slonim, the IBM engineer who leads the Project Debater team, said.

“Off in outer space”

Donvan, the "It's Debatable" host, said that the technology was a much better way to allow the audience to participate in the debate compared to what he and other debate moderators have typically done in the past, which is to simply call on audience members interested in making a comment more or less at random.

"Audiences can be very tricky for me in that I randomly call on people to raise their hands and ask a question of the debaters, and to be honest I have to throw out a large number of the questions because they are repetitive, or they off in outer space, or they are just not well articulated," he said. "This helps with that challenge."

Clea Conner, Intelligence Squared's chief executive officer, said the key point analysis A.I. allowed the company to incorporate views from a much broader and diverse group of people. It does so by enabling people anywhere in the world to submit opinions via a Web page. "This innovation is really helping us understand what a much larger group of people than the four or six hundred that could attend the event live previously, in this case thousands of people, where they really stand on the issue," she said. Slonim said he thought the technology could even be used to make future U.S. Presidential Debates more participatory.

To train an A.I. system to score arguments for quality, IBM had to create a database of 30,000 human-generated positions on a wide variety of topics which were then assessed by small human focus groups of 10 to 15 individuals, Slonim said. The results of this exercise were then used to teach the A.I. what constituted a coherent, strong argument.

The key point analysis feature is a refinement of a technology, which IBM called "speech by crowd," that it unveiled last November in a debate held at the Cambridge Union debating club at Cambridge University in England.

The ability to summarize large amounts of text and pull out key points is an active area of A.I. research with big potential commercial applications. In just the past three months, both OpenAI, the San Francisco research company that was co-founded by Elon Musk and has received funding from Microsoft, and Primer, another startup in San Francisco, have unveiled tools that can summarize long text documents. Researchers at Facebook have also been pursuing this capability too.