As part of our AI in Your Community series, I spoke to Elizabeth Clark, who won the Amazon Alexa Prize for her work with Sounding Board, a social bot. Elizabeth is studying natural language processing and working on tools for collaborative storytelling.
Elizabeth Clark
Tara Chklovski: Tell me a little bit about what you’re working on.
Elizabeth Clark: Very broadly I’m working on natural language processing, so looking at how language and computers interact, and helping computers process language – either written text or speech. More specifically I’ve been looking at collaborative writing systems, which give people support and offer suggestions to them as they write. I’m exploring how we can build models that will generate suggestions that are helpful to people as they try to write, say, a short story. There are different levels to offer help to people as they write. You could point out grammatical errors or spelling mistakes, or you could offer suggestions about structure. The type of suggestions we’re interested in are focused on the actual content for your story.
Our goal is to look at what type of suggestions people want, and determine how we can give them suggestions that are coherent with the story that has come so far, but are still creative and surprising – all to try and spark their creativity as they write. As for what are useful suggestions, we’ve found that it really depends on who is using the system. Different people want different things out of these suggestions. Some people really like silly suggestions, that have these unexpected elements, and they’ll work really hard to try to find a way to work it into their story, embracing it as a challenge…where other people know exactly what they want to write and if the suggestion isn’t in line with that, then they will just delete it and write their own story. There does seem to be a tradeoff between the level of unexpectedness of the suggestion and how coherent it is with what has come before.
Tara Chklovski: Interesting. So how do you know what type of help to offer?
Elizabeth Clark: That’s something that we’re still working on now. The way that we see these systems is to give the author complete control. We’re not trying to replace them as they write; we’re really just trying to support them. So letting them have control over the suggestions and keep as much of the suggestions as they want and to delete as much as they want is one way to handle that, because that might change too over the course of the story. Another possibility is to give them more control over the type of suggestions. So before we provide a suggestion, perhaps we let them choose certain things, like how much they want surprising suggestions versus coherent suggestions.
Tara Chklovski: Can you tell me a little bit more about what the experience of using this tool would be like?
Elizabeth Clark: Right now it begins with a visual prompt, which is an unlabeled cartoon. We ask the writer to write one sentence of a story and once they’ve written one sentence we offer a suggestion for the second sentence. They can go in and edit that suggestion as much as they want before moving on. And then it’s their turn to write a sentence by themselves again, and then we’ll offer a suggestion for the next one. Basically at every other sentence we offer a suggestion that they can then edit as much as they like before moving on.
Tara Chklovski: Have you done an analysis of very popular stories or books like the New York Times Bestsellers, and what elements they share?
Elizabeth Clark: That’s one of the things that we’re really interested in moving forward. Stories are more than just a collection of sentences – there’s a structure that normally goes along with them, a story arc. We’ve looked at characters and the role that they play in stories and story structure. And as we move from this single-sentence suggestion to helping people write in a broader sense, structure becomes more important.
Tara Chklovski: What type of model is behind this?
Elizabeth Clark: Right now we are using a neural network, a sequence-to-sequence model. There aren’t any rules involved, and it’s actually trained on a whole bunch of stories. It generates a suggestion for the next sentence one word at a time. So the sentences it’s providing are novel sentences that it’s offering to the writer.
Tara Chklovski: Cool. So what has been a really surprising story that came out of this?
Elizabeth Clark: I’ve had a lot of fun reading the stories that have come out of the system that people write. And people get really creative. Again, a lot of times it’s the suggestions that maybe are unexpected or surprising that I found people tend to get the most creative with and write the most interesting stories with. We had a story based on an image with three different characters, who had been named in the story, but the computer offered a suggestion with a new character name. It introduces a new character name Furble, and I thought that obviously the author was going to go through and delete the suggestion entirely. But they ended up getting really creative – one of the characters was sick so they named the character’s disease Furble. They just made a little joke about it in the story and moved on, but it was really creative. I think the people that really worked at incorporating these more unexpected elements into their story, those were the ones that usually entertained me the most.
Tara Chklovski: Interesting. So that is an example where the human accepts the computerized suggestion, and it is sort of the amplification or augmentation of creativity.
Elizabeth Clark: Exactly.
Tara Chklovski: That’s awesome. This is an awesome problem to help people write better, more creative stories – what drew you to it?
Elizabeth Clark: I really like this problem because there isn’t just one question involved with it. There are a lot of different directions you can take it in, and the lessons that you can learn from tackling this problem are applicable to a much wider set of problems. It’s nice having a single application or test case to work on, but being able to keep in mind that as you solve these problems there are bigger implications for the work. And I think a really nice starting place for this project is that this is something that pretty much everyone has experience with. Everyone has, at some point in their lives, sat down and tried to write a story, whether it was in a class or just for fun. And it’s also something that people often struggle with.
Tara Chklovski: What inspires you in general, and what inspires you to be in this field?
Elizabeth Clark: One of the first things that got me interested in this field was that I really liked learning languages, linguistics, the grammar of languages. And I’ve also always really loved math. When I was in college I wasn’t sure if I was going to be a Spanish major or a Math major, and then I took my first computer science class just for fun when I was halfway through college and I loved it. And I found out that you could combine all of these pieces together into computational linguistics or natural language processing.I remember one of the first areas of NLP that I became aware of was through things like Google Translate which we had used in the language classes I took when I was younger. And I just started asking questions about how it worked. I was curious to know how you build a system like that, and started thinking about how I would do it, how I would improve it … I really found those questions interesting.
Tara Chklovski: Very cool. I think that gets at the way that computer science infiltrates every field and gets adapted. What have you found really difficult, and how do you overcome some of the challenges that you encounter?
Elizabeth Clark: One difficulty is seeing something that you’ve come up with not work. You have this great idea and you put it into practice, and then nothing comes out of it, or it doesn’t work as well as you thought it would. That can definitely be disheartening, but that doesn’t necessarily mean that the idea that you had wasn’t a good idea. You can spin that and view it instead as an opportunity and recognize that you know more now. You’re closer to finding a solution. It can become a good challenge too, because then you get to go in and ask why it didn’t work and what went wrong. Being able to accept that what you really wanted to work out didn’t and using that to push towards the actual solution is something that I’ve gotten better at, but I’m definitely still working on it.
Tara Chklovski: I think all of us have to keep working on it. How often do you encounter roadblocks like that?
Elizabeth Clark: I think the Amazon Alexa prize chatbot that we worked on was probably the time where that happened the most. For good and for bad, we were live for most of the competition, so we had people interacting with our chatbot all of the time and we had scores and conversation logs which meant we got to see what worked and what didn’t very very quickly. We could try things out constantly and based on what we saw, we were able to learn more about what people actually want out of a socialbot.
Tara Chklovski: That’s fun! How do you think AI will help strengthen societies and communities?
Elizabeth Clark: I think that the potential for AI to support people in a wide variety of tasks is really exciting. We recently published a paper that was about specifically writing with help from a machine. People are really great at creative writing and the more aesthetic elements of writing, and we definitely have not built systems that can write award-winning novels yet or anything like that. But computers are able to perform other types of tasks well. They’re very good at churning out a lot of text really quickly for example, and they have access to a lot of information that they can learn from.
From Elizabeth Clark’s recent paper: Machine-in-the-loop system structure: the loop is initiated with the person providing context and the machine responding with a suggestion. The person always has control over the final result.
So how do we take those strengths that people have – their ability to write creatively, to understand what a good, interesting, creative idea is – and the strengths of a computer and put them together? I think that there are a lot of areas like this that have really great potential for taking the strengths that AI and computers have and using them not to replace people but to support them as they complete tasks.
Tara Chklovski: Totally. And I think we’re very strong believers in that human augmentation part. So what field of AI excites you the most?
Elizabeth Clark: Natural language processing. It’s been exciting to see things like Siri and Alexa because it brings people and technology a little bit closer in the sense that we want to be able to interact with them as naturally as possible. I think there are a lot of exciting problems there to motivate this field of natural language processing.
Tara Chklovski: So the last question: what do you think is the best way for children to learn more about AI?
Elizabeth Clark: I would say asking questions, being curious about the things that you see, and asking about how and why they work. And don’t be afraid to start with a simple solution! Try something simple – you might be surprised at how well it works. Find a problem, ask questions about it, try and come up with a solution and then from there, see what works and what doesn’t and improve it. I think sometimes people think that you need to start with the fanciest system that does everything correctly. But that doesn’t need to be the case. Don’t be afraid to fail. Try something basic. See what it does and you’ll learn from that.
Tara Chklovski: I think that’s awesome advice because the satisfaction of building something yourself is just so great that you keep going deeper and deeper. So I think that’s awesome advice. Thank you so much Elizabeth.