Baidu Beats Google and Microsoft, Wins Championship in AI Language Field

2019-12-30

Baidu has recently developed a new AI language model that has achieved the highest score ever in the recognized General Language Understanding Evaluation (GLUE) benchmark test, surpassing Google and Microsoft.


GLUE consists of nine different tests, including identifying the last name of a person in a sentence and understanding the meaning of pronouns like "it" when there are multiple potential options. The average human score on the GLUE scale is around 87, and Baidu is the first company to score over 90.


Baidu refers to this AI language model as ERNIE (Enhanced Representation from kNowledge IntEgration). Researchers have been working to improve the performance of the GLUE model, so the current standard set by Baidu may soon become outdated. However, what sets Baidu apart is that their learning method seems to be applicable to other languages. Even though the model was developed for interpreting Chinese, the same principles allow it to perform better in English as well. ERNIE follows the development of the BERT language model.


BERT introduced a new standard for language understanding in bidirectional models. Previous language models could only interpret data flowing in one "direction" and considered words appearing before or after the target word as context. BERT enables bidirectional approaches, using both the preceding and following words in a sentence to help understand the meaning of the target word. BERT employs a technique called "masking" for bidirectional analysis, where a word in the sentence is selected and hidden, splitting the possible contexts of that word across the preceding and following context clues.


In English, words are the primary semantic units, and people recognize the meaning of a whole word rather than individual characters. A word can be removed from the context and still retain its meaning, and the meanings of individual characters are almost always the same. In contrast, Chinese relies more on how characters match with other characters to determine meaning. The meaning of a character can vary depending on the surrounding characters.


The Baidu research team essentially adopted the model used by BERT and expanded upon it by hiding strings instead of complete words. They also trained the AI system to distinguish between random strings and meaningful strings, so that the correct strings can be masked. This allows ERNIE to proficiently retrieve information from text documents and perform machine translation. The research team also found that their training method resulted in a model that can better distinguish English phrases compared to many other models. This is because certain word combinations used in English can express different meanings when connected versus used separately, although this occurs rarely.


ERNIE utilizes various other training techniques to optimize performance, including analyzing sentence order and distance when interpreting paragraphs. It also employs a continuous training method that allows ERNIE to train on new data and learn new patterns without forgetting previously acquired knowledge.


Baidu currently uses ERNIE to improve the quality of search results. The latest architecture of ERNIE will be detailed in a paper to be presented at the 2020 Conference on Artificial Intelligence Promotion Association.

RECENT AI NEWS

RECENT AI TOOLS