Talk to Books, an experimental semantic search engine, allows people to ask questions in English and the engine will search more than 100,000 -- about 600 million sentences in aggregate -- to find and highlight the relevant information on the page.
The models driving this experience were trained on a billion conversation-like pairs of sentences, learning to identify what a good response might look like, explains Ray Kurzweil, Google's director of engineering, in a blog post.
Kurzweil introduced the tool Friday.
"You make a statement or ask a question, and the tool finds sentences in books that respond, with no dependence on keyword matching," he wrote. "In a sense, you are talking to the books, getting responses that can help you determine if you’re interested in reading them or not."
For example, ask the website "Can AIs have consciousness?” and the query will return a list of books that include information on that specific question.
Kurzweil and a team at Google Research created the technology. If you type a question into the search engine on the Talk to Books website, it will scan every sentence in 100,000 volumes in Google Books and generate a list of likely responses with the pertinent passage bolded.
Last year, Google used word vector models of language to make improvements to Smart Reply for Gmail. Now, researchers are using the technology to explore other applications, such as Talk to Books and Semantris. Both are found on a website called Semantic Experience.
Semantris, a word association game, allows users to enter a word or phrase. The game ranks all of the words onscreen, scoring them based on how well they respond to what the user has typed.