Even with access to 200 million pages of content, consuming four terabytes of disk storage, IBM’s Watson supercomputer wasn’t able to answer correctly every question during its 2011 Jeopardy appearance. The computer sometimes faltered and flopped – seemingly failing to understand the questions.
Computer Science graduate student Lareina Milambiling’s research explores why that is, and looks to bridge the gap between artificial intelligence and natural language – in the gaming industry.
“It’s a problem of natural language. If I said a sentence that wasn’t grammatically correct, you’d be able to process it – ‘Coffee, now, go.’ A machine would think the sentence order is not proper, and wouldn’t necessarily understand the semantics or context,” Milambiling said.
“IBM’s Watson was a big stepping stone. But ambiguous sentences confused it.”
Milambiling completed a master’s degree in Linguistics in 2013. Among the first cohort of Western’s Continuing Studies Diploma in Computer Science program this year, she received a Natural Sciences and Engineering Research Council of Canada (NSERC) grant to fund her artificial intelligence work as part of her masters in Computer Science, which she started this fall.
“I was trying to figure out what to do with my (Linguistics) research, because it’s very theoretical – semantics and syntax – and how to draw that out into a job. And computational linguistics is a big thing – it’s why I did the one-year diploma at Continuing Studies,” she said.
Tackling something like artificial intelligence by way of video games, Milambiling is scaling down a larger challenge and forging an accessible route to its solution.
“In games like World of Warcraft and Minecraft, when you’re talking to artificial intelligence in the game, you’re clearly talking to a machine. You have to select what you’re trying to say,” she explained.
But game developers would like to see the game process what humans are actually doing, Milambiling added.
“Then AI (artificial intelligence) agents in the game can react based on what players are doing. We’re trying to achieve human level intelligence in the game – it’s a controlled environment, so it’s easier to do,” she continued.
IBM’s Watson was tasked with answering questions in any imaginable context, which is why it had trouble coming up with correct answers. Tackling the issue of natural language and artificial intelligence in a video game makes it more accessible, Milambiling explained.
“In a game environment, context is very limited. If you put a machine out in the real world with thousands of different contextual factors changing, it would have no idea how to change the semantics,” she said.
“I’m hoping to develop my project in Minecraft – it’s a popular game and you have a lot of people interacting to get more data.”
Computer Science professor Mike Katchabaw credits Milambiling’s solid background in Linguistics with positioning her perfectly.
“This work has entertainment applications in the gaming industry, as well as non-entertainment applications,” said Katchabaw, who supervises Milambiling. “It is very exciting work, and it is great to have someone with a dual background like Lareina here for this sort of thing.
“She is uniquely qualified for study in this problem space.”