Computers bypassed men in verbal IQ test Bashny.Net

More than a hundred years ago, the German psychologist William Stern (William Stern) proposed a test of human intelligence assessment, which was named test IQ. Since then, the IQ test was quite widespread as a standard method to estimate the intelligence of children at school entry, as well as for the evaluation of adult candidates for the job.

IQ tests generally contain three types of questions: 1) questions of logic where you need to recognize a pattern in the sequence of images; 2) mathematical questions, where you define a pattern in a sequence of numbers; 3) verbal tasks based on analogies and classifications as synonyms and antonyms.

Researchers from Microsoft Research division in Beijing, together with colleagues from the University of Science and Technology of China developed the technology of AI capable of solving problems of the third type of the above ( scientific article < / a>).

Computers could never have a good understanding and solve problems formulated in verbal form. At least, they did much worse people. Microsoft Research Development changes things. Their program is based on a system of deep learning (deep learning), the first time exceeded the average results shown by people in solving word problems from test IQ.

In previous years, scientists have used the technique of mining data to analyze large amounts of text to find specific connections between words. In particular, this technique allows you to compile a dictionary of statistical indicators of how often certain words are located nearby. This allows you to determine the relationship of words to each other.

As a result, each word in such a system is seen as a vector in a multidimensional parameter space. Such a system of vectors can be handled by mathematical methods: compare them, add, subtract one from the other, as conventional vectors. For example, it is possible a similar equation: "King - man + woman = Queen».

This approach has proven effective. For example, Google uses the data mining system of automatic translation of texts by comparing the vectors of words in different languages.

But in the case of verbal IQ tests the problem is complicated because there is one word can have several meanings. Compilers of tests specially made so that complicate the task.

A group of researchers from Microsoft Research division found a solution for this problem by using the same data-Mining: their program determines which words most often found each word in the array of texts, and then determines the possible meaning of the word, on the basis of the information received. This is done by calculating the vectors of the proposals received. For the first program in a matrix of frequency of occurrence of words, and then based on the corpus (Wikipedia article) for each occurrence of the word indicates a vector with it other words.





Scientists say that the program shows a better result than most people. The survey was conducted online people Mechanical Turk.



According to the table, its result is about halfway between the results, which show average results bachelors and master's degree holders.

Source: geektimes.ru/post/252196/