|
又到了每周一次的 Nature Podcast 时间了！欢迎收听本周由Nick Petrić Howe和Shamini Bundell带来的一周科学故事，本期播客片段里讨论了可以和人类辩论的AI。欢迎前往iTunes或你喜欢的其他播客平台下载完整版，随时随地收听一周科研新鲜事。 音频文本： Host: Shamini Bundell First up on the show, it’s time for a bit of a debate. Interviewer: Nick Petrić Howe In 2019, a historic debate occurred. Project Debater Greetings Harish. I have heard you hold the world record in debate competition wins against humans, but I suspect you’ve never debated a machine. Welcome to the future. Interviewer: Nick Petrić Howe That right there is the voice of an artificial intelligence called Project Debater, who came up with the statement and the whole debate herself. In fact, she has successfully debated humans live several times. On the podcast, we’ve talked about a few different AIs – ones that can play board games, some that can play video games, and AIs like Watson who competed on Jeopardy. But for the principal investigator of Project Debater from IBM, Noam Slonim, these challenges aren’t tough enough. Interviewee: Noam Slonim Although it is clear that all these grand challenges were extremely instrumental to the development of artificial intelligence, these board games still lie in what we refer to as the 'comfort zone’ of artificial intelligence. Interviewer: Nick Petrić Howe Things like board games and video games have a clear 'win’ state that an AI can try hundreds of techniques to achieve. Not so in the case of debate. Here, winners are tricky to identify, but now IBM have developed such a system by honing a few different AI technologies, such as argument mining and understanding of human language, and getting each of these different AI components to work together all whilst squaring off against an opponent. Project Debater, who looks a little bit like the monolith from 2001: A Space Odyssey, has been able to face off against top-tier human debaters in front of a live audience, such as the debate I played earlier which took place in California against Harish Natarajan, a world-class debater. Despite this challenging environment, Project Debater has done pretty well for herself. Interviewee: Noam Slonim In the three public debates that we had, we lost one, and we were able to win another one and it was nearly a tie in the third one, and in addition I think it was interesting to note that in all the debates that we had we also asked the audience another question: which side better enriched your knowledge during the debate? And in all of the debates, Project Debater obtained clearly better scores than the human opponent on this question, which was perhaps expected but still nice to see. Interviewer: Nick Petrić Howe This week in Nature, Noam and his team are publishing a comprehensive rundown of the technology involved in Project Debater. But before we get into how the system works, it’s worth quickly running through how the debates it participated in were laid out. To start, there’s a motion – what are we arguing about here? For example, should pre-schools be subsidised? Next, there’s opening remarks – four minutes of argumentation about why your side is right. Then after hearing what the opponent has to say, there’s four minutes of rebuttal. And finally, two minutes of closing remarks. So, at the start, Project Debater is given a position to argue for and then has 15 minutes to prepare the opening remarks. Project Debater For decades, research has demonstrated that high-quality preschool is one of the best investments of public dollars, resulting in children who fare better on tests and have more successful lives than those without the same access. Interviewer: Nick Petrić Howe So, here are some of the remarks that Project Debater is coming up with. But how is she doing it? Well, she draws on a wealth of human argumentation. Interviewee: Noam Slonim First of all, it has a large collection of around 400 million user articles from LexisNexis, nearly 10 billion sentences. And when the debate starts, the system is using various AI components to detect short pieces of text that should satisfy three criteria. They should be relevant to the topic. They should be argumentative in nature – that is they should argue something about the topic, not just be relevant. And finally, they should support our side of the debate. Then after finding these short pieces of text, the system is using the other AI components, like text clustering and et cetera, in order to glue these short pieces together into a meaningful narrative. Interviewer: Nick Petrić Howe Now, finding these bits of text is a huge challenge. Identifying the bits that meet those criteria are part of what Noam’s team have been working on since the project started in 2012. Project Debater largely achieves this through having lots of data and ranking how relevant those bits of sentences are. She then groups these sentences into topics using Wikipedia to help. After all that, Project Debater has to assemble those sentences into a coherent argument. Again, not an easy task. It’s hard to say exactly what it is that makes a debate compelling. And then there’s rebuttal. Project Debater For starters, I sometimes listen to opponents and wonder: what do they want? Would they prefer poor people on their doorsteps begging for money? Interviewee: Noam Slonim You need somehow to respond to the arguments of the opposition, and this starts by really understanding the words articulated by the human debater, and for that purpose, we used Watson’s speech recognition capabilities out of the box, but of course you need to go beyond the box. You need somehow to understand the gist of the human speech and the main claims being raised, and to that end we developed several techniques that typically rely on the same principle of trying to anticipate in advance what kinds of arguments the opposition might use and then listen to determine whether indeed the opposition was making these claims and then respond accordingly. Interviewer: Nick Petrić Howe Altogether, as Noam mentioned, Project Debater was pretty good at debating. Compared with other AIs, Project Debater was ranked more highly by audiences and close to that of a human expert, but not without some limitations. Sometimes, she found it difficult to make an argument flow like a real human does and, not too surprisingly, she kind of argued like machine, with lots of facts and figures and not as much emotion as a human debater. But whether or not Project Debater is better or worse than a human, it’s a big step for the field. Finding arguments in human written text – so-called argument mining – and language generation, well, these are tricky tasks for an AI, as AI researcher Elena Cabrio explains. Interviewee: Elena Cabrio Even for humans, for which debating is among the primary cognitive activities, I mean, with every day, so even when we debate, we need to apply a wide range of language understanding and language generation capabilities. So, for a machine, being able actually to address all these tasks at the same time in an automated way is actually a big improvement in the field. Interviewer: Nick Petrić Howe For Elena, who was not involved in this project, the ability of AIs to search through vast amounts of text and fine-caught arguments could help with an ever-growing problem in the modern world – information overload. Interviewee: Elena Cabrio The growing of the web, the increasing number of texts and data that are published every day, have actually highlighted a need to process such data in an automated way, to be able to identify, structure and summarise this huge amount information. People like us are more and more exposed to information – online newspapers, blogs, online debate platforms, social networks – so argument mining has actually the potential to help us with that because it provides the techniques to sift through this ever increasing amount of data and provide us with the relevant evidence items that we can find in them. Interviewer: Nick Petrić Howe Noam feels similarly, and sees Project Debater as more of a collaborator with humans, potentially helping them find arguments to assist their own debates or speeches. For both Elena and Noam, the next steps for this kind of research are to try and improve these AIs abilities to understand language as humans use it. What makes a good argument? What is convincing? Why is this compelling? But for now, Project Debater has given an insight into what this sort of technology could look like. Project Debater Thanks for this final opportunity to speak out in this debate and thanks, Harish Natarajan. One might say that this conversation can serve no purpose anymore, but I feel differently. Interviewer: Nick Petrić Howe When you saw Project Debater up there debating with people, what did that feel like? Interviewee: Noam Slonim I was proud. Interviewer: Nick Petrić Howe That was Noam Slonim from IBM Research. You also heard from Elena Cabrio from Université Côte d’Azur in France. If you feel the matter is still up for debate then you can find a link to Noam’s paper in the show notes. 《自然》论文： An autonomous debating system |
|
|
