Ů��С��Ƶ

Soon you could be chatting with your computer about the morning news. An AI has learned to read and answer questions about a news article with unprecedented accuracy.

Creating AI systems that can learn in the background from humanity’s existing stores of information is one of the big goals of computer science. “Computers don’t have the kind of general knowledge and common sense of how the world works [from reading] about things in novels or watch[ing] sitcoms,” says at Stanford University.

To get a step closer to this, last year, Google’s DeepMind team used articles from the Daily Mail website and CNN to help to read and understand a short story. The team used the bulleted summaries at the top of these articles to create simple interpretive questions that trained the algorithm to search for key points.

Now a group led by Manning has designed an algorithm that beat DeepMind’s results by an impressive 10 per cent on the CNN articles and 8 per cent for Daily Mail stories. It scored 70 per cent overall.

The improvement came through streamlining the DeepMind model.  “Some of the stuff they had just causes needless complications,” says Manning. “You get rid of that and the numbers go up.”

Design trade-off

“It makes sense,” says of Carnegie Mellon University in Pittsburgh. “Making something more complicated doesn’t make it better.”

There’s a trade-off in AI design: if an algorithm is complex, it’s more powerful, but to perform well it needs more data to learn from, says Frederking. Simpler AI can train quickly with smaller amounts of data.

Manning says there’s not much more a computer can learn from this particular data set. To prepare the hundreds of thousands of articles for AI readers, DeepMind used a program to go through them and assign the same label to nouns and the pronouns that reference them. But this program inevitably confused some pronouns. Fresh nouns and pronouns labelled in a new data set would be needed to keep AI reading improving.

The advantage of using Daily Mail and CNN articles was that there’s so many of them, says at the University of Illinois at Urbana-Champaign. The more texts the algorithm learns from, the smarter it becomes. It will be tricky to find or create another large set of texts that come with ready-made questions, she says.

Before we turn artificial intelligence loose to gather knowledge from the world’s texts, there’s a few challenges to overcome, says Frederking. “It’s hard to keep these things on track and figure out what information to keep and what to throw away,” he says. “If you’re not careful, your AI system will think Obama was born in Kenya.”

Reference:

Read more: Computer with human-like learning will program itself; Super-literate software reads and comprehends better than humans