How to give a computer common sense
Computer scientists have built a system that can make simple guesses — for example, what happens after “Gary stacks kindling and logs and drops some matches”. Machine-learning techniques, such as neural networks, can make statistical guesses after absorbing large amounts of data, but they are notoriously lacking common sense. Other attempts relied on knowledge bases containing millions of rules and sentences handpicked by humans. COMET combines the two approaches, using a knowledge base to train a neural network. Lead researcher Yejin Choi says she was surprised that no one had tried this approach before. “It’s almost as if nobody bothered because they were so sure this would never work.” COMET’s guesses are still statistics-based, but at least they’re often right.
Quanta | 14 min read
No hay comentarios:
Publicar un comentario