Google's AI Is Able To Summarize Texts Without Copying Fragments Of The Original Document

Text summarizing is something AI has made great progress at, as stated by a recently published report by Microsoft. This is a great piece of news for those who feel tired of reading long messages every day.

Google, as one of the largest tech companies in the world, wants its fair share of the development too. A partnership between one of its Google Brain teams and experts at Imperial College London have made Pegasus, its own machine learning system. According to the search engine giant, its AI has managed to complete 12 summarizing tasks with subjects covering stories, patents, emails, news, science, legislative bills, and instructions.

According to the researchers, what the AI is able to do is abstractive summarization, which produces linguistically precise copy without copying fragments of the original document.

The team trains the AI by feeding it documents with important sentences hidden. What the machine learning system has to do is to fill in the gaps by searching the web as well as the corpus the researchers compiled.

After the experiments, the experts picked out the model with the best performance they have. Pegasus has 568 million parameters, was trained with 750GB of texts from 350 million web pages.

The AI is able to produce summaries with a high level of coherence and fluency, without countermeasures to lessen disfluencies. In another setting with a more limited data set, one with only 100 example articles. Pegasus was able to produce summaries at a similar quality as another model with a more extended data set of 20,000 to 200,000 articles.