The charity that maintains Wikipedia has developed AI that can spot contradictory claims in articles and alert the human editors who write and edit the collective encyclopedia.
Wikipedia was launched in 2001 and the English language version now has more than 6 million articles written almost entirely by volunteers. While that huge, unpaid workforce is its biggest strength, it is also a crucial weakness, and errors can be introduced either accidentally or deliberately.
Cheng Hsu at National Cheng Kung University in Taiwan and colleagues believe that automation using artificial intelligence can provide a way to spot errors in the vast cloud of content and flag them for human review.
Advertisement
Working with the Wikimedia Foundation, which operates Wikipedia, the team created an AI that can detect contradictions within articles. For instance, the Wikipedia article on musician Tyler Acord was found to claim, in the same article, that he was born in a town called Lakewood and also in a town called Renton. The software was able to spot that these facts cannot both be true and raise this as an error.
AIs known as neural networks require large amounts of training data and examples of the type of problem they are intended to solve. This was provided by Wikipedia鈥檚 thousands of human editors, who can flag an article as having contradictions.
The team scoured all English Wikipedia articles up to March 2020 for these contradiction warnings and found 2321. The researchers then looked at subsequent edits to those pages to see which contradictions had later been solved by human editors, giving them 1105 examples of contradictions and sensible solutions. The neural network was trained on this collection of examples and taught to spot contradictions on its own.
When the neural network was trained with 80 per cent of that data and tested on the remaining 20 per cent, it was found to have an accuracy of up to 65 per cent. This could dramatically increase the number of errors caught and fixed.
The Wikimedia Foundation already uses autonomous software bots to correct misspellings and to start new articles from collections of geographical or statistical data. It has also used AI to scour hundreds of thousands of daily edits for malicious vandalism.
Reference:
