Baidu defeats Microsoft and Google in teaching AI to understand human language

Chinese search giant Baidu easily defeated two US tech giants Microsoft and Google in a contest of natural language processing of the world's leading AI models.

The General Language Understanding Evaluation (GLUE) is the standard used to train, evaluate and analyze AI-based natural language understanding systems. Organizations and businesses developing artificial language processing models based on artificial intelligence can assess their models using the GLUE benchmark.

Until recently, Microsoft's MT-DNN-SMART model was still the leading name in the global GLUE rankings, followed by Google's T3 model. However, this ranking has just been a big disturbance because in Baidu's new natural language processing contest, Baidu's ERNIE model (Enhanced Representation through kNowledge IntEgration) defeated both MT-DNN-SMART and T3 with a record score of 90.1.

Initially, Baidu's model was developed to understand the Chinese language, but researchers also soon realized the prospect of excellent understanding of the model. Baidu's ERNIE training method is heavily inspired by Google's 'old' language processing model training tool called BERT. Accordingly, 15% of words in each series will be hidden at each training, the AI ​​task tries to predict the hidden words as closely as possible based on the context. The Baidu team illustrated this technique on their Github site with examples cited from the Harry Potter novel series. As a result, ERNIE was able to extrapolate the relationship between Harry Potter and JK Rowling by analyzing tacit knowledge of words and entities, to conclude that Harry Potter was a novel set written by JK Rowling.

'Although understanding language is still a difficult challenge, our results for the ERNIE project have shown that language training models are continually trained and multi-tasking as one direction. promising for NLP research, 'said the Baidu team.

Baidu is now pushing the ERNIE model application for practical tasks. For example, Baidu is currently using the ERNIE model for its automated query answering feature, which improves 16% of user satisfaction with search results.