Maarten Sap

Maarten Sap

As part of the AI in your Community series, I recently spoke with Maarten Sap, a PhD student at the University of Washington. Maarten is interested in natural language processing, and social science applications of AI. Maarten is also the 2017 winner of the Alexa Prize, an Amazon competition to further conversational artificial intelligence. Maarten and his team created Sounding Board, a socialbot that finds interesting content to talk about with users while trying to understand who the user is.

Tara Chklovski: Tell me more about the Alexa Prize!

Maarten Sap: The Alexa Prize was a collaborative effort. I’d been working in conversational modeling, including therapy conversation modeling, when our team got together to work on a project for the Alexa Prize. The challenge was to converse about anything and everything, starting with an open-ended prompt, and there wasn’t really structure or a goal to it; it was really confusing. We went into it asking “what does it mean to have a conversation?”

We thought about how conversations are a way to exchange information, and then Hao Fang, our team lead, mentioned using different Reddit threads to find really interesting facts, which then led us to wonder what topics people are actually interested in. And from there we asked ourselves how we could best present that to people, and how to follow up from one topic to the next so that the conversation feels coherent.

There was a lot of trial and error, which was kind of interesting because the trial and error happened at a large scale that we’d never operated on before. It was in people’s homes! Everyone who has an Echo can talk to our bot. If you have an Amazon Echo and you’re in the US you can just say “let’s chat” to Alexa and talk to one of the finalists.

The scale of it was unlike anything we’d ever done, and it really forced us to develop something that was much more polished than a prototype. We were talking to adults and children and older people of all different demographics. We really had to think about who we’re talking to, which I thought was really relevant because oftentimes in research people don’t necessarily think about all the populations that their tools could affect.

TC: Interesting. How did you identify who you were talking to?

MS: We actually had no idea because the prize rules kept us from knowing. But Amazon would hint at the customers they had, and we thought about all the possible people that we could be talking to. We had a clever sort of icebreaker that we incorporated. We wanted to get to know the person we were talking to – not just demographics like age, but what their personality was like. So, we ended up asking them actual personality questions like, “are you usually the life of the party?” which was a way of turning social psychology theories into a tiny part of a conversation. We interspersed those questions with simpler things like whether someone would rather have vanilla or chocolate ice cream, and all of that was a way to get the conversation started. It was kind of like speed dating.

TC: I was just going to say that. That’s what human beings do, right?

MS: Exactly. And after greetings, people didn’t really know what to talk about and so we thought it would help to know each other really quickly and play some icebreaker things. And that was really a strong part of our chatbot in the end.

TC: Very cool. I’m also very intrigued by your other research project around movies. Are you still working on that?

MS: Yes, we’re still working on that. One of the things that I’m particularly interested in is how we make AI more socially aware – how we take into account majority class versus minority classes, gender differences, gender bias and so on.

For this movie bias project, we were interested in exploring how to measure whether there’s gender bias in character portrayals in movies. We can computationalize that by looking at the actions that characters take, and specifically whether or not they give characters a lot of agency and power or whether they imply negative agency and limited power.

TC: How do you do that?

MS: We do this with “connotation frames” which are annotations based on a verb. You can determine if the verb gives the subject or the object more power. We annotated approximately 2,000 verbs, and then we could look at what verbs are used to describe which characters and count the times the characters are being empowered or not.

For instance when the verb “implore” is used in a sentence like “he implored the tribunal for mercy” there’s an inherent power dynamic implied by the word “implore.” He has less power than the tribunal for instance, right? Whereas if he is demanding mercy from the tribunal then he has much more power over it because he’s demanding instead of begging or asking.

We selected these verbs based on the 2,000 most frequently used verbs in our set of movie scripts. We set up a simple annotation tool for humans to use where we just asked, for instance, “if x implores y, does x have more power or does y have more power?” We used Amazon and Mechanical Turk, which is a crowdsourcing platform.

TC: What did you find?

MS: What we found was not very surprising. Female characters in movies tend to have less agency. They’re less decisive and they make fewer decisions about their destiny, and they also are more submissive or in less powerful positions. And then the opposite applies for male characters.

We also noticed that there was a trend of female characters that take a lot of powerful actions but there tend to be only one or two of them, and there tend to be very few other women that had power, and they wouldn’t interact with each other.

Sap's work examines gender bias in movies based on verb data.

Sap’s work examines gender bias in movies based on verb data.

TC: Interesting. What advice would you to children and families around trying to find a problem that they can tackle using AI technologies? We find that a big stumbling block for them is figuring out where to start and understanding how to go about narrowing down the field. How did you come across the movie project?

MS: The movie bias project actually started when we were trying to detect “mansplaining” and then realized that condescension is a highly content-sensitive thing to tackle, and there wasn’t much we can do from the natural language processing side of things. So, then we moved on to what we could measure.

One key thing that was always present was the larger purpose of what we want to do with AI – making the world a better place. I know this probably sounds really cheesy but trying to further our personal ideals and try to get AI to do that, that was always something that we used as motivation.

For instance, the next step of the movie bias project will be exploring how to actively correct for the negative portrayals of women in movies instead of just measuare it. Adding those personal ideals and letting them guide us is definitely a way that helps us solve interesting problems.

TC: Right. I think staying true to yourself and figuring that out is the key. So, a couple of questions about yourself: what inspires you?

MS: I’m inspired by social science, psychology, and social psychology, and specifically tackling societal commentary using AI and Natural Language Processing techniques. My angle has always been exploring how we potentially correct for societal problems by using NLP tools.

TC: I think that’s a huge area, and I feel like that’s more and more important that people who work with technology have to understand social psychology.

On the positive side, how do you think AI can help strengthen society and communities?

MS: I think the accessibility side of AI definitely is going to help. I also think AI could strengthen society by more aware of the inherent biases in society and trying to mitigate them, because AI has been helpful in making quick decisions. I think where we could really strengthen that is by adding  a more socially-aware component to that AI.

TC: Right. What field or aspect of AI excites you, looking forward?

I’m more interested in language, like natural language processing. I think language is an interesting concept because a behavior but it’s also a way of communicating. And I think understanding that and understanding everything that goes on in my head while I’m speaking is interesting. Having systems understand text and being able to reproduce human language with the complexities and idiosyncrasies that people have is something that really excites me.

TC: I think it’s something that is accessible to everyone because language is something that binds us together. One last question – what do you think the best way is for children to learn more about AI?

Not science fiction!  As much as movies have inspired a lot of current products (the Amazon Echo was basically the computer from Star Trek), I think that a lot of science fiction has been something of a doomsday scenario version of AI.

I think understanding basic computing can get children pretty excited. Basic computing is like Lego blocks but it’s logical Legos; you’re putting different logic blocks together to create a small program. From there the jump to AI is not a big jump because instead of adding the logic manually, you’re helping the model learn the best logic from data.

I remember I was really excited as a kid when I was playing with Lego robots and got the robot to follow directions so it would go right if you pressed the right button and left if you pressed the left button. And really, setting up any appliance was also what got me into computer science in general. I was always really excited when my mom would get a new printer or something like that.

I think for children, just getting to play with something that’s super harmless and fun and colorful is the best way to do it, and that’s why the Lego stuff is definitely really cool.

TC: You’re totally spot on that’s why we’ve curious about that AIY Voice Kit because it’s cheap and it allows you to make these rule-based conversations and it gives you the building blocks, and it gives you a taste of automation.

MS: And I think it’s really cool that people like you are spending the time to educate families and children because not everyone gets an opportunity to play with these things at a young age. A lot of people come to college and that’s the first time they have computer science experience, while at the same time there’s a lot of privileged children who’ve been playing with this stuff their whole lives because they have access to it.

TC: Totally. And I think a big part of it is actually trying to set them up for success than it is to sort of give them a boost, which is why we’re talking to researchers like you, Maarten. Thank you so much for doing this.