Chinese technology company Baidu outperformed Microsoft and Google in the General Language Understanding Evaluation, a benchmark for artificial intelligence (AI)'s comprehension of human language.
Researchers built the Enhanced Representation through kNowledge IntEgration (ERNIE) model specifically for the Chinese language, using a method that also makes the model better understand English.
The researchers built on the Bidirectional Encoder Representations from Transformers (BERT) model, which predicts and interprets the meaning of words by considering context before and after words simultaneously. BERT randomly hides 15% of words in a given passage of text and attempts to predict those words from the remaining ones, and Baidu modified this technique for ERNIE, hiding strings of characters rather than single characters; ERNIE also differentiates between meaningful and random strings to mask the right character combinations.
Baidu uses ERNIE to produce more applicable search results, eliminate duplicate stories in its news feed, and make AI-facilitated responses to requests more accurate.
From MIT Technology Review
View Full Article
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA