This AI learnt language by playing games

Parlez-vous Monkey Island 2?

Researchers at MIT have put together a computer system that can complete sections of text adventure games without any prior knowledge of how language works - suggesting that it can learn the meanings of words.

Regina Barzilay, Tejas Kulkarni and Karthik Narasimhan wanted to design a system that could make inferences about the syntax of language - that could work out the difference between "you are hurt" and "you are not hurt" and what it means without being taught.

Negations and Conjunctions

To do so, they built their own text adventure game that relied heavily on negations like this, as well as conjunctions like "but", "and" and "or". They also asked the developers of Evennia, a game-creation toolkit, to build a test game too.

The computer system they created was tasked with trying to get through the two games, and its performance was compared with two other techniques for natural-language processing. On both, their system outperformed the rival techniques - though it was stumped when it was asked to compare two different descriptions of an object.

"When you play these games, every interaction is through text," said Narasimhan. "It's not like a console with buttons. So you really need to understand the text to play these games, and you also have more variability in the types of actions you can take.

The details of the experiments were published on Narahimhan's website.

Image credit: Jose-Luis Olivares/MIT // CC BY-NC-ND 2.0