Google teaches AI to write Wikipedia articles

Google Inc. has long been working in the field of artificial intelligence . All this is necessary for the company both for personal needs (improvement of the work of various services), and for providing new opportunities to its customers. Now Google has announced a new function of its AI, which was taught to “read” a variety of texts and form a Wikipedia-style article based on the read results.

AI takes materials from the Internet, and analyzes them, and then creates something like a summary. According to the results of their work, employees of the corporation wrote an article indicating many technical issues. The main idea of ​​this article - reading analysis is a very difficult task for the machine. Before Google, a similar task was performed by other companies, but it never worked out for them - most of the similar projects failed.

But there are those who succeeded. For example, the company Salesforce was able to train its neural network to collect information and analyze it according to a given algorithm. This is similar to what Google did, and the results are also available as an article. The result can be called satisfactory - the neural network was coped with the common task, but the text generated by the computer was not very good. The proposals, for the most part, were simple and short.

Also, the computer was difficult to maintain the semantic structure of content, with which a person copes without much difficulty. As for Google, this company turned out a little better. The offers that the computer gives out are longer than those of SalesForce and look more natural. So far, the machine can only work with English texts. Below is an example with two different notes. Left - the one that was written by man. On the right is the “typewritten” material, information for which the neural network was looking for on different resources.



The text that the computer writes is somewhat hard to read due to, for example, the lack of capital letters at the beginning of new sentences. In addition, most of the proposals are somewhat heavy. However, in general, this text reads well. In this particular case, not all the important information from the source code got into the final version, so the example itself cannot be considered very indicative. Nevertheless, it is clear that the neural network is really capable of much - at the moment it is the best version of what a computer can produce after analyzing the original text.

The system searches for information on a given topic on different sites and generates a list of the ten most relevant pages containing the necessary data. As a result, an integral article is formed with links - about the same format as that of Wikipedia. By the way, the system created by Google, self-learning, so that each page read and created helps the computer to improve.

By the way, the computer deliberately shortens long sentences to make it even more similar to the style adopted for the articles of the Wikipedia electronic encyclopedia. Thus, the computer does not write everything by itself - it only forms its own sentences, reducing the original ones, from which you can throw out part of the words, without distorting the meaning of the read material.

The process can not be called too fast, but the result is quite acceptable. According to the experts who formed the algorithm, in this whole project, the “bottleneck” is the selection of separate text fragments to form a new article. All this work is not done very quickly, so the algorithm has where to grow and improve. Representatives of the corporation said that now the software and hardware for such work is not too perfect, so the company plans to improve its technology further.

The main drawback of the current system is that the computer takes the information pages for writing an article on the subject, based on their popularity (visits, reading time, etc.). And if the information placed on one of the pages is not too accurate, then the summary, which the computer creates, will also contain errors and inaccuracies. Nevertheless, the company hopes that in the near future it will be able to show an even more perfect process of forming texts.