Microsoft has developed an AI that can read the text and answer questions about what has been read.
Recently it became known that the Microsoft Research Asia team was able to achieve significant success in the development of AI. Using the Stanford Question Answering Dataset dataset, known among researchers as SQuAD , scientists have created a weak form of AI that can read the text and answer questions about the read no worse than a person (even slightly better).
Using the SQuAD scale, researchers assessed the success of their offspring. As it turned out, the AI scored 82.650 points, while the result of a person with a similar set of questions and answers is 82,304. As it turned out, Microsoft is not the first company to achieve such a success. Before it, about the same number of points scored by AI from Alibaba, then improving its result after reaching Microsoft. Now the software of both companies is in the first place of the SQuAD rating.
The SQuAD test itself is a questionnaire containing about 100,000 questions. They are based on 500 excerpts from articles on Wikipedia. In the process of passing the test questions, you must not only read but also listen. It is believed that this test is one of the most accurate tools for measuring the abilities of artificial intelligence.
Among other issues in the Stanford test, there are exercises related to the history of the British series Doctor Who. The computer needs to answer the name of the main character's spacecraft. Companies such as Facebook, Tencent and Samsung have previously presented their own products.
Previously, Alibaba experts have already expressed their satisfaction with what is happening. Thus, Luo Xi, a leading specialist in natural speech processing, says that the results achieved are an important event in the history of creating artificial intelligence. “This technology can be applied step by step in various areas, from customer service and museum guides to medical online consultations. Human involvement in such processes can be minimized, ”he said.
According to the Financial Times journalists, the fact that the Chinese and Americans are engaged in the same issues, and with approximately equal results, indicates a clear opposition of both countries in such a field as the development of artificial intelligence. The Chinese authorities have announced the development of AI as a priority in the country. China plans to become a leader in this area between 2020 and 2030. China, in principle, has all the chances to achieve the planned due to the fact that the country has a huge number of network users and incredibly large data arrays on the basis of which it is possible to train AI.
As for the United States, here, while the government cautiously touches upon the topic of AI. In particular, the country so far only formulates the essence of the term, preparing regulatory laws. Here, a lot of attention is paid to the ethical side of the issue, which is not so important in China. In China, Chinese companies are engaged in the development of AI, which are ready to share their data for the sake of a common result. By the way, Baidu management has already announced that the data packages used in the course of their work, as well as the results of the research, will be provided to the scientific community.
“Such tests are definitely useful indicators of how far the work on AI can go. True, the main benefit of AI for a person can manifest itself only if the AI acts in harmony with people, ” saidMicrosoft representative Andrew Pickup.
The corporation uses the results of its achievements in its own products. For example, early versions of AI trained with Stanford texts are already used in the Bing search service. Now the company is developing a new version of the product that will be able to answer several related questions. For example, after answering the question: “What year was the German Prime Minister born?” A computer, according to Microsoft experts, should be able to answer the question: “In which city was it born?”. And the machine must understand who and what is being discussed in the new question, giving an adequate answer. We still have to wait for new achievements.