We’ve seen artificial intelligence beat humans at games so often we’ve resigned ourselves to the fact that machines armed with bundles of data can easily trump a human in a video or board game.
But an AI project coming out of IBM is equal parts frightening and fascinating. That project is called Project Debater.
Since 2012, researchers in Israel have been working on an AI program that is able to competently debate a topic against a human.
To achieve this the researchers had to push the boundaries of artificial intelligence.
The researchers describe the complexity of understanding language on something called the continuum of complexity. For instance, something such as interpreting the sentiment of a sentence can be easily achieved by a good AI. But translating language as it’s spoken and summarising what has been said is something humans are still far better at.
A debate is somewhere between those two extremes on the continuum but is closer to the complex end.
“Debating represents a primary cognitive activity of the human mind, requiring the simultaneous application of a wide arsenal of language understanding and language generation capabilities, many of which have only been partially studied from a computational perspective (as separate tasks), and certainly not in a holistic manner. Therefore an autonomous debating system seems to lie beyond the reach of previous language research endeavours,” reads an article outlining the researcher’s finding in Nature.
Project Debater is made up of four main modules.
The first of these is argument mining, which is pretty much force feeding the AI a corpus of information. In this case it was 400 million news articles from the LexisNexis 2011 – 2018 corpus. These articles are then index according to the words the mention, the entities mentioned, “the Wikipedia concepts they refer to”, and the dictionary definition of words.
When given a debate motion, Project Debater relies on this index to formulate its argument.
The next module is the argument knowledge base or AKB.
This module finds common threads in debates. For instance, if the topic of the debate was banning a substance, it could draw from data relating to an underground market for instance.
“Texts in the AKB contain principled arguments, counter-arguments, and commonplace examples that may be relevant for a wide range of topics. These texts are authored manually – or extracted automatically and then manually edited – and are grouped together into thematic classes,” explain the researchers.
When a debate motion is provided to the AI it is able to use the AKB to find information most relevant to the topic.
From there the next module comes into play – Argument rebuttal.
This is where Project Debater formulates a list of claims its opponent might present. This module sees the AKB and argument mining modules working together. The AI also uses arguments extracted from iDebate just in case the debate is being repeated.
Finally, the debate construction module sees the AI formulating the final debate. Irrelevant or unnecessary information is removed and the debate is fine tuned and arranged into clusters which are then all evaluated. The debate itself is done via an expressive text-to-speech solution “developed to suit argumentative content”. We hope that text-to-speech solution also contains sighs, just in case.
So how well does it work?
Well, thankfully we can see it with our own eyes. Project Debater and debating champion Harish Natarajan squared off in a live debate in 2019 that you can check out below.
That is by far the scariest video we’ve seen in a long time – and we are avid true-crime enthusiasts.
The good news is that we aren’t a point where machines can reason better than a human can but with projects like this and how well the appear to be performing, it might not be too long before AI and humans are really battling it out in debates.
You can read more about Project Debater here and we highly recommend you give it a read.