Microsoft DeBERTa

Main article: Speech technologies: on the way from recognition to understanding

2021: The AI model of Microsoft exceeded result of the person in the test for SuperGLUE natural language understanding

On January 15, 2021 the Microsoft company reported that the algorithm of a natural language understanding (Natural Language Understanding, NLU) Microsoft DeBERTa exceeded human opportunities in one of the most difficult tests for similar algorithms SuperGLUE. For January, 2020 the model wins first place in rating with an indicator in 90.3 while mean value of human opportunities makes 89.8 points.

The rating of algorithms on passing of the SuperGLUE test

According to the company, the SuperGLUE test includes a number of tasks which are developed for assessment of capability of AI models to distinguish and understand a natural language, for example, to give the correct answer to a question based on the read paragraph, to define whether the multiple-valued word in a certain context, etc. is correctly used. The test was developed by group of researchers in 2019. When SuperGLUE was provided, the gap between the most effective model and indicators of the person in the table of leaders made nearly 20 points.

To achieve the current result in 90.3 points, DeBERTa received large-scale updating of architecture: now it consists of 48 layers and has 1.5 billion parameters. Microsoft will make public model and its source code. Besides, DeBERTa will be integrated into the next version of the Tyyuringovy Microsoft Turing model (Turing NLRv4). Tyyuringovy models are used in such products of Microsoft as Bing, Office, Dynamics and Azure Cognitive Services to optimize, for example, interaction with chat-bots, providing recommendations and answers to questions, search, automation of a customer support, creation of content and solution of many other tasks to advantage of hundreds of millions of users.

Architecture of the DeBERTa model

Unlike other models, DeBERTa considers not only words meanings, but also their positions and roles. For example, in the offer of "a new store opened beside the new mall" (engl. "the new shop opened near shopping center") it can understand that relatives on contextual store value ("shop") and "mall" ("shopping center") play different syntax roles (a subject is "store" here). Moreover, it is capable to define dependence of words from each other. For example, DeBERTa understands that the dependence between the words "deep" and "learning" is much stronger when they stand nearby (term "deep learning") than when they occur in different offers.

In spite of the fact that the DeBERTa model exceeded human indicators in the SuperGLUE test, it does not mean that the AI model reached the level of the person in a natural language understanding. In difference from machines, people well are able to use knowledge which is earlier received at accomplishment of different tasks, for the solution of others – it is called composition generalization (engl. compositional generalization). Therefore, despite promising results of DeBERTa in the test, it is necessary to continue researches to develop this skill at model.

Microsoft actively works on optimization of artificial intelligence technologies. So, in October, 2020 the AI algorithm for automatic generation of signatures to images which in many cases are more exact, than written by people was provided. It will allow to make products and Microsoft services even more inclusive and available to the bigger number of users. First of all, the automatic description helps people with visual impairment during the work with documents or web pages and also in general allows to get access to contents of any image, for example, by search or preparation of the presentations.