Artificial intelligence from Microsoft has outperformed humans in interpreting human voice

In the SuperGLUE test for natural language comprehension, Microsoft’s DeBERTa AI model set a milestone. She scored 90.3 points, and he scored 89.8 points. In less than two years, Microsoft has obtained those performance and aims to further boost its AI.

Superintelligence artificials:


In the SuperGLUE benchmark, a measure of natural language comprehension, Microsoft’s DeBERTa artificial intelligence (AI) has exceeded human skills, earning 90.3 points, while a person’s metrics are at the level of 89.8 points. The difference is, in other words, tiny, but still there.

A Natural Language Understanding (NLU) algorithm generated by Microsoft is DeBERTa (Decoding-enhanced BERT with disentangled attention).

The SuperGLUE exam consists of eight activities relating to recognizing the text’s meaning as well as its meaning and interactions of cause and effect in it. Among other items, during his DeBERTa (Decoding-enhanced BERT with disentangled attention) passage, he had to address questions about the paragraph he was reading, and also to explain if a certain word was correctly used in a certain context, for which the language offers many definitions.

The AI read, for instance, the expression “the child became immune to the disease” (in the original – the child became immune to the disease). The question “what is the reason for this?” was asked (what is the reason for this?), and DeBERTa gave two answers to it – “he avoided exposure to the disease” and “he received the disease vaccine”… Although selecting the correct answer is simple for a person, it is challenging for an AI model. The model must consider the causal association between the premise and two (or more) possible answers to get a correct answer.

The evolution of artificial intelligence at Microsoft

In 2019, the SuperGLUE test itself was created by a team of researchers, and the smartest artificial intelligence was almost 20 points behind human results at the time of its debut. This suggests that it took Microsoft less than two years for its DeBERTa to not only keep up with people in natural language comprehension, but to exceed them.

The AI model’s entire architecture had to be profoundly changed by its experts, and now it comprises 48 layers and 1.5 billion different parameters. The organization will make the model itself and its source code available – all the required details will be released on the GitHub resource that has been owned by Microsoft since June 2018.

DeBERTa has an important distinction from other AI models, in addition to the 1.5 billion parameters, which have played a part in reaching 90.3 points in SuperGLUE. By implication, this AI takes not only the meanings of words into consideration, but also their locations and functions in the sentence. “For example, he is very capable of recognizing that the word “market” and the term “shopping center” (mall) play numerous syntactic roles in the phrase “a new store opened next to the new mall”

Microsoft experts have instructed their brainchild to decide that words rely on each other. For starters, DeBERTa knows that when they are side by side (the expression “deep learning”), the relationship between the terms “deep” and “learning” is much greater than when they occur in separate sentences.

The computer revolt is also very far behind.


The fact that DeBERTa has managed to outperform humans in the SuperGLUE benchmark does not mean that the degree of human comprehension of natural language has been surpassed by artificial intelligence in any way.

Microsoft demonstrated that a human is good at applying his expertise gained during his life while performing such requirements and assignments, unlike computers, to solve new tasks assigned to him. Compositional generalization”compositional generalization”
The corporation has a proposal to use DeBERTa’s capability. This AI model is to be implemented into the next iteration of Microsoft’s Turing model (Turing NLRv4).

Turing models are used by Microsoft throughout its entire product portfolio, including the Bing search engine, the office suite, and Azure Cognitive Services, the Office cloud provider. They are used to increase engagement with chatbots, provide feedback and answers to requests, scan, automate customer service, build content, and more, among other things.

Not united by Microsoft
Microsoft’s AI model really dominated humans in the SuperGLUE survey, taking first place, but individuals were not ranked second, but just third in the ranking. Silver was awarded to another model of artificial intelligence, T5 + Meena, by Google.

The score is 90.2 points, which means that it is just 0.1 points lower than that of DeBERTa. Google did not reveal information about its AI model at the time of publishing of the test results in SuperGLUE, writes the VentureBeat portal, and because of that it managed to achieve such high success in this benchmark.