A new reading AI that beat humans for the first time in one of the world’s most difficult reading comprehension tests could have a significant impact on various industries, says the academic director of NextAI.
“This type of technology could allow a system to rapidly find the correct answer to a question posed by a customer,” explains Graham Taylor, who is also a machine learning and software engineering assistant professor from the University of Guelph. “Any industry that relies on humans reading or search large databases of text could potentially benefit. This includes the legal, finance, and medical industries.”
Another specific example, says Taylor, is converting free-form, unstructured documents into an organized document.
“Consider a set of legal documents like a lease, which are all written in slightly different ways, but we want to extract the same specific pieces of information from each. A good reading comprehension system should be able to solve this task,” he explains.
Alibaba’s Institute of Data Science and Technologies (iDST) recently announced its deep neural network model scored 82.44 in the Stanford Question Answering Dataset, beating the human score of 82.30.
“We look forward to sharing our model-building methodology with the wider community and exporting the technology to our clients in the near future,” said Luo Si, iDST’s chief scientist for Natural Language Processing, in a statement.
Taylor sees the technology being applied to technical support systems. It will be interesting, he adds, to see the reading AI develop to the point where it can produce an output where there are countless, or an infinite amount of correct answers.
“However, this poses other challenges around coming up with the right error metrics to train and evaluate such systems,” he says.