Apr 4, 2019 - Technology

An AI-written science book

Book with binary code coming out of it.

Illustration: Sarah Grillo/Axios

At the start of the first chapter of a thick new academic tome about lithium-ion batteries, where information about the author would usually go, five words hint that something's different: "This book was machine-generated."

Why it matters: AI is helping to speed up science, unlock impossible problems and dig researchers out from under information overload. Automatic summarization is a big remaining challenge that, if solved, would accelerate discovery by focusing researchers on the most pressing problems in their fields.

What's happening: The book, which summarizes peer-reviewed research papers about lithium-ion batteries, is the first machine-written volume from Springer Nature — but the publishing giant says more are on the way.

  • The following 232 pages are a dry, technical read. But it's entirely intelligible — valuable, even, for a scientist trying to catch up to the vanguard of battery research.
  • Most work on AI writers has focused on fiction. It's not easy, as we've reported, to get a computer to generate good sentences and to fill them with actual facts.
  • Past nonfiction efforts — like several that have tried to create textbooks from Wikipedia articles or other sources — often stumble over readability issues.

Details: The machine-generated summary is made up of intelligible sentences, but it's anything but a pleasant read. It's pocked with citations — a result of the computer's inability to understand themes and concepts — and many sections are a paragraph long, built around just one research paper.

  • "The holy grail in this space is a system that can read multiple texts, understand the texts, integrate those texts, synthesize the ideas, and then turn that into generated language," says Kristian Hammond, a Northwestern professor and co-founder of Narrative Science, a language AI company.
  • "We're not even in the ballpark of that. We're nothing."

How it works: The "prototype" volume was written by an algorithm that the publisher created with researchers from Goethe University in Frankfurt, Germany.

  • The algorithm — "Beta Writer," as it's credited on the cover — picks out relevant research from Springer's abundant archives, groups it by similarity, and arranges it into chapters and sections. Then, it summarizes the work, footnoting as it goes.
  • Check out the free e-book, with its detailed intro about Beta Writer.

But, but, but: By using only research published by Springer Nature, the book is missing out on the vast body of peer-reviewed work published elsewhere — like the prestigious Science journal, for example.

Go deeper