Baidu Beats Google With New AI Language Training Technique
Chinese tech giant Baidu has created a new way of teaching language to AIs, according to a report in TechnologyReview. The new method offers better results than the ones used by Google and Microsoft, beating out both companies in the General Language and Understanding Evaluation (GLUE) competition.
BERT and ERNIE
Baidu’s new model is called Enhanced Representation through kNowledge IntEgration, or ERNIE. It was named after the Sesame Street character because Google named the former champion model the Bidirectional Encoder Representations from Transformers or BERT. To take the crown from Google, ERNIE had to outperform its rival in the nine different language tests of GLUE. The tests the ability of an AI to understand context, proper names, and other aspects of a language. People are able to score an average of 87% on the 100-point test. ERNIE’s average is more than 90%, the first AI to ever do so.
What made BERT revolutionary is in the bi-directional part of the name. Earlier language models could use the word before or after a target word to determine context. BERT can use all of the words in a sentence to decipher the target words meaning using a technical approach called masking, where some of the words are hidden and the AI attempts to predict them using context clues. ERNIE extends the masking to strings of characters. That’s critical for understanding Chinese where individual characters change meaning based on the ones around them, but the same technique is useful for training an AI in English and other languages as well, especially for informal speech and colloquialisms.
That isn’t the only way ERNIE trains AIs to understand language. Continuous training, sentence order, and other elements are drilled into the program as well, but the character masking is the most important improvement. ERNIE’s reign may not last long, as researchers continue to improve AI learning capabilities. But, the generalization of Baidu’s creation is significant and likely to inform all future AI language training systems.
Smarter Teaching, Better Voice Assistants
Baidu has already begun using ERNIE to improve how well its Xiao Du AI assistant responds to people speaking to it. A better language interpretation engine is going to be a huge boost for Baidu and any other voice assistant developer who starts applying it. Understanding what people are saying is a central part of the competition between voice assistant developers. It’s part of why the companies pay attention to things like the IQ test for voice assistants run by Loup Ventures, which Google Assistant has been winning of late, though Amazon’s Alexa is closing the gap. On the other hand, a recent report by Perficient saw a universal decrease in the ability of voice assistants to answer questions accurately.
Xiao Du was not tested in either case, however, so there’s no direct comparison of the Chinese voice assistant and its American rivals. The extra edge of a better language learning engine could give Baidu an edge in the near future, although Baidu’s presence in the space globally is already rapidly rising. In a year it went from a small source of smart speakers to the number two seller globally and number one in China, with an assistant in more than 400 million devices.