Q&A with Matty Mariansky

How can chatbots help with education?

“There is a university in the United States that awards scholarships, where it was discovered that out of 100 people who start the process, and who submit the first form, very few reach the last form. It’s a process that includes many stages, and a lot of people just drop out in the middle. A chatbot is a creature that’s really good at reminding you, encouraging you, teaching you, walking you step by step, when the process is very, very long. It tells you, ‘OK, look, the next step we have to take is to send this form. Yesterday we said you’d send it and you haven’t done it yet. Are you sending it? OK, it’s sent and seems to have been accepted, but it has not reached the person who handles such requests. Call her and sort it out.’ It knows how to walk you through processes that might have worn you down. It’s never exhausted. For example, in Europe it is very difficult to switch cellphone carrier. It requires lots and lots of stages. People simply do not switch; they never move to a different provider. But they’ve been able to build a chatbot that helps you with this. You’ve taken a step, and a month later you’ve forgotten about it. The bot tells you, ‘Remember we took this step a month ago? It’s time to take the next step.'”

So You’re saying the advantage of chatbots in this area is that they can lead people on a path that is difficult, exhausting, bureaucratic, multi-stage – they both direct you when you are ready to give up and forget about it, and they save the other side the need to send you reminders. In this sense, maybe a chatbot is suitable for teaching tasks that suffer from the same problems, such as finishing exercises, doing homework, and repeating certain material.

“I want to be able to wake you up in the middle of the night, and have you know how to answer my question. A chatbot can wake you up in the middle of the night and see if you know the answer,” he laughs. “Another thing chatbots are good at – in theory, anyway, as it’s very difficult to develop – is helping you to adopt good habits. For example, I talked to someone who has a chronic illness, and he has to record daily what he eats and how he feels, to try to find out what foods are causing him stomach pain. When he changes his diet, it takes three months to notice the difference in the abdominal pain. One cannot follow these changes over a long period. One actually has no immediate incentive; he gets no reward for writing down every day if his stomach hurts or his head hurts. A chatbot can nag you, but in a way that motivates you. First of all, it can try to break the habit. Instead of telling you every day, ‘Good morning, today we will measure your blood pressure,’ it can say it in a different way every time; for example, by changing the time when it is done, or showing you the progress you’ve made – ‘You’re fantastic, you’re the best, we’re on our way, we’ve already done half, we’ve already halfway through’ – all sorts of things that are very difficult for you to do because you’re not persistent; there’s no one to nag you and make you do it while giving feedback.”

Let’s apply this to teaching – if you study math or a new language, your progress depends on the previous step you took, but sometimes it’s hard to see the progress. A chatbot can be used for this purpose?

“It doesn’t give up on you. If you haven’t made progress, it can stay with you in level 2 until you’re good enough to proceed to level 3, and it will then praise you – ‘There, we moved up a level.’ It can adapt to the way you answer or the way you use it. It doesn’t have to have been built in advance with a fixed program. It can be built to be adaptive.”

Adaptive, as in suitable for a specific user and not one size fits all?

“Exactly. A glove for your own hand.”

For the full Magazine >> Click Here

And you also reduce the burden of the teachers, who do not need to sit in at each student’s individual program?

“They can get a general report – Yossi is doing well, Danny is on the second stage, and Ronnie didn’t answer at all.”

This means that the chatbot manages the studying for the student, but also for the teacher?

“A teacher has to give a lesson to an entire class, and all the students listen to the exact same thing. A chatbot is like The Sorcerer’s Apprentice, where the broom is split into 200 smaller brooms, and each one of them fetches water. It is the same with a chatbot – the teacher can split himself into 200 teachers, and the same process works for each student personally, but they don’t all listen to the same lecture and not everyone progresses at the same rate. But they all have the teacher in front of them simultaneously.”

Will the way we assess student success change as well? If each student learns at his own pace, and gets his own curriculum, then it is not presumed that eventually all of them can take the same test.

“Possibly. The grading can also be adaptive, perhaps according to the individual student’s starting point. It can be tailor-made for each student according to his needs. There may be a project that has four types of tasks – assembly, a math problem, drawing a picture, and singing a song – and maybe everyone starts at the same point, but the bot recognizes the strength of each student and steers him to a different part of the exercise, where it sees that the student has a chance of succeeding and completing it. That could be interesting – to make a program that not only brings you from point A to Z where you only change the pace, but that makes the way you get from A to Z completely different for each student. One student builds a model of a volcano because he is good with his hands; another analyzes the physics of the volcano – each student will do what he is best at, that he is comfortable doing, and complete the project from a point that’s good for him.”

Could a chatbot also help with educational orientation? It’ll work with a student for a while and then tell him which area is the most suitable for him.

“The big promise of chatbots is that they have neither buttons nor menus – the idea is that you simply talk to them and they understand what you want. And when that promise is fulfilled it will be great, because people have no better way of communicating what they want than by talking. Think what it would be like if you could talk to everything – you’d tell the door, ‘I want to get to the other room,’ and it will open; you would tell the faucet, ‘I’d like the water a little warmer,’ and the water will be warmer when you wash your hands. When the chatbots’ language understanding reaches its peak maturity, you’ll be able to work in this way. You could tell the chatbot, ‘I don’t understand this exercise. It’s just this part I don’t understand – just explain it to me; maybe we can do something else right now because I’m finding it a bit boring.’ You won’t need to think about how you should phrase it, which words you should be using – which is what you have to do today.

“Currently, you either have to use very specific wording, or the chatbot’s programmer has built in some kind of menu that you choose from, which is a bit like calling the phone company and being told, ‘dial 1 for sales, 2 for whatever.’ This promise cannot be fulfilled because picking from a menu is not that interesting, and when you want to tell the chatbot what you want and it doesn’t understand, you’re thinking, ‘OK, I’m talking to an idiot, I don’t feel it can teach me anything – it doesn’t even know how to speak.’ And all the beautiful things we’re talking about crash into the wall of users’ frustration. You really have to sit down and think about how to adjust the bot so that it will be able to elegantly evade things it doesn’t understand, know how to steer the conversation back to its comfort zone, and understand as many sentences as possible that a student might use. The programmer needs to think about these things in advance, as well as fix the bot while it’s working. That means that when you start to run a chatbot in a real classroom, and see what the students are saying and where the bot is failing, you need to fix it immediately. With the first, second, and third students it’ll fail, but with the fourth student it will already understand the type of request being made and know how to respond.”

That is to say, it will really learn, the way it should function – as a learning machine?

“At the moment, most of the learning is things we fix. We are the ones looking at the negative feedback loop and correcting it where necessary. So I can go in every day and see all sorts of things that people said to the bot which it did not understand, appearing really stupid sometimes. And I try to fix the specific thing I see that day, because it’s very difficult to anticipate all the things that users might tell it. The fact that you have a device that you can speak to in Hebrew and it understands everything – that’s its big drawback. Because it allegedly understands everything, and there is no specific button to click on, people may tell it things like ‘Let’s take a break, I want to eat a sandwich. Make me a sandwich.’ You never imagined that anyone would tell a bot ‘make me a sandwich.’ So you can decide that if they said something like that, the bot should say, ‘I did not understand what you said. Maybe we’ll do it later. Let’s go back to studying.’ You do not have to think specifically about a sandwich. If you told the bot to say, ‘I have no idea what the answer to that question is,’ then when the student says, ‘Tell me, robot, why do you exist at all?’ the bot will answer, ‘I have no idea what the answer to that question is.’ Once we identify the mechanism under the hood, it’s very easy to manipulate, send it to the entire class, and find the nastiest question according to the bot’s answer. You can confuse it and make up any question so that it comes out as if it’s been answered in the way you intended.

“So long as understanding our language hasn’t reached the level you see in Space Odyssey, this promise is very difficult to fulfill with regard to certain things. It does work if you customize it very carefully, considering all the possible scenarios, formulating the questions in a way that guides it as to what kind of answer you expect. If one develops a chatbot, it’s not a matter of sending and forgetting. One should expect to launch it and then keep improving it every single day, until it reaches a certain standard, because it’s really hard to guess on the first day in which direction people will take it, as it has no buttons and nothing in it is pre-structured, so that you can just answer A or B or C. And if you did just do A-B-C you’ve ruined the whole experience. It’s no longer a chatbot – it’s a phone menu.”

For the full Magazine >> Click Here