🗞 Better fact-checking for fake news
WHAT: Bias is nothing new in AI, but there is a fascinating irony when fact checking models are trained on biased data. That is the case for many fact checking models that were trained on the popular Fact Extraction and VERification (FEVER) dataset. Fortunately, MIT professor Regina Barzilay created a debiased FEVER dataset and developed a new fact checking algorithm that performs better than the others across every metric.
WHO: Regina Barzilay is a professor at MIT and a member of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). She’s interested in NLP, and its intersection with chemistry & oncology.
WHY SHOULD I CARE: ~45% of Americans get their news from social media, which is a perfect breeding ground for fake news. Furthermore, Gartner predicts that if current fake news trends persist, “a majority of people in the developed world will see more false than true information.” Developing robust AI-driven fact checking systems can prevent this dark prediction from becoming a reality.
TL;DR: Many of today’s fact-checking algorithms were trained on a popular but biased dataset. Researchers have now debiased this dataset and developed the best-performing fact-checking algorithm to date.
READ THE PAPER HERE