Communications of the ACM
Learning Language By Playing Games
Massachusetts Institute of Technology (MIT) researchers have developed a computer system that learns how to play a text-based computer game, called Evennia, with no prior assumptions about how language works. Although the system cannot complete the entire game, it can complete sections of it, suggesting it can discover the meanings of words during its training.
The researchers evaluated the system by comparing its performance to that of two others, which use variants of a technique standard in the field of natural-language processing. The basic technique is called "bag of words," in which a machine-learning algorithm bases its outputs on the co-occurrence of words, while the variation, called the "bag of bigrams," looks for the co-occurrence of two-word units. When playing the Evennia game, the MIT system outperformed systems based on both bags of words and bags of bigrams techniques.
The new system used deep learning and relied on two performance criteria: completion of a task in the Evennia game and maximization of a score that factored in several player attributes tracked by the game. On both measures, the deep-learning system outperformed bags of words and bags of bigrams. "I think . . . the general area of mapping natural language to actions is an interesting and important area," says Stanford University professor Percy Liang.
From MIT News
View Full Article
Abstracts Copyright © 2015 Information Inc., Bethesda, Maryland, USA
No entries found