Artificial intelligence can now summarize entire books like people

OpenAI is able to generate “sensible” summaries of entire books thanks to improvements in its algorithm.

Advertisements

The company specialized in Artificial Intelligence (AI) OpenAI  has taught its latest algorithm to summarize complete books in a reasonable way, which in some cases obtain a rating comparable to the average of summaries written by human beings.

The new system solves the difficulties that automatic algorithms usually have to summarize text by combining two elements: human-reinforced learning and what is known as decomposition of recursive or repetitive tasks, as reported by OpenAI through a statement.

As it does?

On the one hand, the automatic mechanism is developed with the help of people, who provide their opinion and assess the quality of summaries of a part of the text produced by Artificial Intelligence .

On the other hand, to accommodate longer, whole books, OpenAI breaks the task down into several parts and first summarizes shorter parts of the text, which it then combines to produce a complete summary of the entire book.

This approach allows for greater ease in human evaluations, since it is not necessary for them to spend hours reading the entire book but only fragments, as well as having greater traceability of the summary process and addressing texts without limit in length.

In this way, the company has claimed that its Artificial Intelligence is capable of generating “sensible” summaries of entire books, using Alice in Wonderland as an example. “Sometimes [the agorithm] is capable of matching the average quality of human-written abstracts,” their authors have come to assert.

In this way, 5% of the abstracts of a complete OpenAI book obtain a score of six points out of a total of seven, the same mark as those written by people. 15 percent, meanwhile, get five points out of seven.

OpenAI’s model uses the latest version of its language-based Artificial Intelligence algorithm , GPT-3, released in summer 2020, capable of supporting 175 billion parameters.

This tool can complete a text from an initial sentence, as well as adapt it to different writing styles, with legal vocabulary or as if it were an article for a magazine.