Photo illustration: Axios Visuals
Artificial intelligence researchers have tried unsuccessfully for decades to give machines the common sense needed to converse with humans and seamlessly navigate our always-changing world. Last month, Paul Allen announced he is investing another $125 million into his Allen Institute for Artificial Intelligence (AI2) in a renewed effort to solve one of the field's grand challenges.
Axios spoke with Yejin Choi, an AI researcher from the University of Washington and AI2 who studies how machines process and generate language. She talked about how they're defining common sense, their approach to the problem and how it's connected to bias.
How do you define common sense?
"Common sense is fairly trivial everyday knowledge that we have about people and about the world. It's knowledge about how the world works — how people think, what motivates them, how they act, and why they do what they do."
"Imagine there's a robot in your household in the future, and you want to store leftover pie in a container. The robot should pick a container that's large enough to store that pie, and today that spatial reasoning relative to physical properties of different objects in the world and how you interact with them are not quite well represented in these system models."
Why common sense poses a challenge to machines:
"We have a world model in our mind when we do daily operations. AI systems today, despite tremendous advancement in recent years, they are not very good at generalizing out of pure example, so they tend to be very, very task specific, and very domain specific."
"A machine translation system may seem like it understands some language enough to translate into another language, but actually there's not that much understanding happening, per se, because that syntax knowledge cannot be reused for making very trivial small talk with a human, for example."
"We have this commonsense knowledge without our parents or teachers having to enumerate all of it one by one. Nobody told us that elephants are usually bigger than butterflies, however we can reason about it. You ask me that question, I can think about it, and I can answer that question even though I've never seen that statement explicitly written anywhere."
"We're taking a similar approach. It may be possible that we can learn to answer these sort of questions, even including those that we've never seen before. That's fundamentally the ability that AI systems needs to have — dealing with unknowns and previously unseen situations."
What data is needed to make common sense models?
"There is a paper called Verb Physics, and in that work the dataset is basically a combination of a lot of natural language documents — a huge corpus of how people use language and from that we look for patterns. For example, what kind of things do I throw? What kind of things do I enter into? I enter my house. I exit my house. And, that sort of implies that my house must be bigger than me for me to enter into and exit from."
"So, we can infer different action dynamics, preconditions, and post conditions — all those different physical objects, for me to do some action involving them."
"The short term goal is to develop a common sense benchmark dataset. Then, the ultimate goal is to acquire knowledge that's good enough to do well in that benchmark dataset. That's step one."
The issue of AI and human bias:
"We showed [in a study of movie scripts last year] how women in movies carry much less power compared to men. It's the kind of actions that they do and the kind of language they use when they speak. Men usually fight and they do stuff, they save the world. Women, on the other hand, they tend to wait, they are being watched, and they look pretty. What they do tends to be pretty passive."
"It's one of my passions to develop AI technology that can detect all these biases in humans and also, ideally, be able to correct them in the future."
How is bias connected to common sense?
"These are connected in that the way bias is coming across, often times can be inclusive or implied. Current models are much better at understanding what's explicitly stated, but less good at anticipating what's not said. It's good to be catching some of the explicit biases, but it's important to also detect all of the implied ones because that still influences us. The ability to read between the lines ultimately is what requires common sense, so that's the connection between the two."