The AI Behind Duolingo • Dr. Burr Settles • Duocon 2021
发布时间 2022-07-06 08:17:44 来源
[♪ OUTRO MUSIC PLAYING
]
请欣赏结尾音乐。
Outro music playing意思是播放结尾音乐。
[♪
Hello, my name is Burr Settles. I'm a research director here at Doolingo, and I'm very excited to talk to you today about some of the ways that we're using artificial intelligence to improve the ways that we teach through Doolingo. Now, to start things off, I want to talk a little bit about our mission, which is to develop the best education in the world and make it universally available. And we're not kidding when we say universally available.
大家好，我是Burr Settles，是Doolingo研究总监。今天我非常兴奋地想跟大家谈一谈我们是如何运用人工智能来提高Doolingo的教学方式的。首先，我想说一下我们的使命，即开发世界上最好的教育，并让其普遍可及。我们说“普遍可及”并不是开玩笑。
I've been with Doolingo for more than eight years, and from day one, we were committed to building a free app that does not put educational content behind any kind of paywall. And that's also why we use technologies like the Internet and mobile devices to deliver our lessons to as many people as possible. But we care equally about it being the best education.
我已经和Doolingo在一起超过八年了，从一开始我们就致力于建立一个不会在任何付费墙后面放置教育内容的免费应用程序。这也是为什么我们使用像互联网和移动设备这样的技术来向尽可能多的人提供我们的课程。但我们同样关心它成为最好的教育。
Now, to a lot of people, the best education comes from having a private one-on-one tutor or a really, really, really great teacher. And while that may be true in many cases, unfortunately, not everybody in the world has access to a one-on-one private tutor or a great teacher. So we believe that artificial intelligence is the best way to connect the dots and really achieve our mission. And to be clear, we're not interested in using AI to replace great teachers.
现在，对很多人来说，最好的教育方法是有一个私人一对一的导师或一个非常好的老师。虽然在很多情况下这是对的，但不幸的是，世界上并不是每个人都能接触到一对一的私人导师或一个好老师。因此，我们相信人工智能是连接教育和实现我们目标的最佳方式。明确地说，我们并不打算使用人工智能来取代很棒的教师。
On the contrary, through platforms like Doolingo for Schools, we want to use technology to enable teachers to do an even better job in the classroom. But until everybody in the world has access to a really great teacher, we think AI is the best way to scale that kind of experience to as many people as possible.
相反地，通过像学校 Doognio 这样的平台，我们想利用技术让教师在课堂上做得更好。但直到全世界的人都能够接触到一位真正出色的教师之前，我们认为 AI 是将这种体验推广到尽可能多的人的最佳方式。
Now, when you think about a really great teacher, I claim that they have three properties. One is that they know the material really well. Great teachers are domain experts in their chosen field. Number two, they know how to make that material engaging to make it exciting and interesting and motivate you to keep learning. And number three, and perhaps most importantly, they're able to get inside your head.
现在，当你想到一位非常优秀的老师时，我认为他们具备三种特点。第一，他们非常了解教材内容。出色的老师是他们所选择领域的专家。第二，他们知道如何使得学习材料变得吸引人，让它变得有趣、有趣，并激励你不断学习。第三，或许最重要的是，他们能够深入你的内心。
So great teacher is because of the one-on-one time that they spend with you and because of the time that they spent with hundreds of other students in the past, they know how to understand the things that you know, the things that you don't know, the things that come easily and the things you struggle with, the things that you've probably forgotten by now and it's time to review.
优秀的教师之所以如此出色，是因为他们与你的一对一时间以及他们与过去数百名学生相处的时间，他们知道如何理解你所知道的东西，你不知道的东西，容易掌握的东西以及你遇到困难的东西，也知道你现在可能已经忘记的东西需要重新复习。
So what we've done at Duolingo is we've taken each of these three properties and turned them into research programs for our artificial intelligence efforts. So for example, when it comes to the material, we use machine learning and natural language processing to create tools that assist our world-class content developers in auditing and improving our courses.
在Duolingo，我们将这三个特点分别转化为我们人工智能研究项目的基础。比如说，我们利用机器学习和自然语言处理来创造工具，协助我们杰出的内容开发者对我们的课程进行审核和改进，以保证材料的质量。
So this can be anything from analyzing the vocabulary and grammar content of our lessons to looking at scripts for Duolingo stories or Duolingo podcasts or audio lessons to make sure that we're hitting a target level of proficiency be it beginner, intermediate or advanced. When it comes to making the material engaging, we're using state-of-the-art speech technologies to create custom voices for our Duolingo world characters, which is super cool and you're also going to be hearing more about that today. And believe it or not, whenever you get a push notification to remind you to do your daily Duolingo lesson,
这项工作包括对我们课程的词汇和语法内容进行分析，查看Duolingo故事、Duolingo播客或音频课程的剧本，确保我们达到了初学者、中级或高级的目标水平。在制作有吸引力的教材方面，我们使用尖端的语音技术为Duolingo世界人物创建自定义声音，这非常酷，今天你们还会听到更多相关信息。信不信由你，每当你收到Duolingo的推送通知提醒你完成每日的课程时，我们的团队都在背后忙碌着为你设计更好的学习体验。
well, it turns out there's an AI algorithm that shows which out of hundreds of possible messages that we could send you at this precise moment to maximize the chances that you'll want to go and do a lesson on Duolingo. Now, those are just a few examples, but I'm going to spend most of our time today talking about how we use artificial intelligence to get inside your head and personalize your learning experience.
嗯，事实证明有一种AI算法能够展示数百种可能发送给您的信息，以最大化您想去Duolingo上学课的机会。这只是其中的几个例子，但今天我会大多数时间讲述我们如何利用人工智能深入了解您的想法，个性化您的学习体验。
And in particular, I'm going to focus on a new system we're really excited about that we affectionately call bird brain. Now, before I tell you what bird brain is, let me tell you a little bit about why we built it.
我要着重介绍我们非常兴奋的新系统——“鸟脑”，并告诉你我们建造它的背景。在此之前，我先要告诉你我们为什么要建造它。
An educational psychology, there's a concept known as the zone of proximal development and it goes something like this. Whenever you're learning a new skill or ability, there's a sphere of things that you understand that come easily, that you've already mastered and you can do without receiving any kind of help.
在教育心理学中，有一个叫做"近发展区间"的概念，它的意思是：当你正在学习一项新技能或能力时，有一些事情是你已经掌握并且能够轻松完成的，不需要任何帮助的。这个领域内的事情属于你的舒适区，而你希望进一步拓展自己的能力的区域则称为近发展区间。
Now, just outside of that sphere is the zone of proximal development. Now, this is at the frontier of what you're able to do. You can still maybe do those tasks or learn those things, but you might need extra time or extra help. And beyond that, there are things that you're just not ready for. They're too hard at this moment.
现在，在那个范围之外，是发展潜能区。这个区域正是你能力的边界。你可能仍然能完成那些任务或学习那些东西，但可能需要额外的时间或帮助。在那之外，有些事情你就还没有准备好。在此时，它们太难了。
Now, there's lots of evidence to suggest that you're most challenged and motivated when you're learning at the edge of your abilities, or if you spend most of your time here in the zone of proximal development. Now, what that means for Duolingo is that whether you realize it or not, when you go and do a lesson, there are hundreds or thousands of possible exercises that we could try to cram into that five-minute chunk of time.
现在有很多证据表明，在你学习能力的边缘或者在你大部分时间处于近发展区时，你会面临最大的挑战和动力。对于Duolingo来说，这意味着无论你是否意识到，当你去做一课时，我们可以尝试将数百或数千个可能的练习塞入这五分钟的时间里。
And so, we want to maximize that time by filling it with lessons that are in your zone of proximal development for you personally. Things that are not too easy are not too hard, but are just right to keep you challenged and motivated. So, you might be wondering how can AI help get you in the zone? Well, that's where bird brain comes in, and here's how it works.
因此，我们希望通过填充适合你个人能力发展区域的课程，最大化利用你的时间。这些课程既不太简单也不太难，恰好能够让你保持挑战和动力。你可能会想知道人工智能如何帮助你进入这个区域？那就是鸟脑发挥作用的地方，以下是其工作原理。
So, for every learner and every exercise, bird brain can look at these two things and make a prediction that this learner has an 81% chance, for example, of getting this exercise correct. Now, some other exercise that might be more basic, it'll predict 98%. Like, this is super easy. But some other more challenging exercise, bird brain might make a prediction like 45%. This is probably something that it's too soon to try to put into your life.
因此，对于每个学习者和每个练习，鸟脑可以观察这两个因素，并预测这个学习者有81％的机会正确完成这个练习，例如。现在，对于一些更基础的练习，它可以预测98％。就像这非常容易。但是，对于一些更具挑战性的练习，鸟脑可能会做出类似45％的预测。这可能是一些你还没有准备好尝试将其纳入你的生活中的东西。
But, that's all for the same learner, some other learner who might be further along on their language learning journey might find this particular exercise a little bit easier, like 79%. And this might be in or closer to their personal zone in proximal development.
然而，对于同一个学习者来说，这个特定的练习可能会有些难度，可达到79%的成功率。但是，对于其他进阶的语言学习者来说，这个练习可能会更容易些，偏近于他们的学习能力发展区。
So bird brain is able to do this because our algorithms learn to make accurate predictions for more than five billion exercises every single week. And this works because each of our individual learners have their own unique profile of the sorts of things they tend to get right and tend to get wrong.
鸟脑之所以能够做到这一点，是因为我们的算法每周能够学习超过50亿个练习，从而能够进行准确的预测。而这种方法之所以有效，是因为每个学习者都有自己独特的学习档案，可以了解他们擅长和不擅长的知识点。
So think about this. The next time you go and do a lesson on Duolingo, you're not only learning something new for yourself, but you're teaching bird brain a little bit about your personal abilities as well as the difficulties of all the exercises that we give you. And this in turn helps us to improve and personalize the learning experience for hundreds of millions of learners all around the world, which is very exciting for us.
想想看，下一次你在Duolingo上学习一课时，不仅是为自己学到了新知识，同时也向我们那笨鸟头传授了一些关于你个人能力以及这些练习的难度的信息。这有助于我们改进和个性化地为全球数亿学习者提供学习体验，这对我们来说非常令人兴奋。
Now, you might be wondering how do we know it's effective. One way we can do this is to look to see whether or not bird brain is making accurate predictions. Now, this chart shows the prediction quality of bird brain over the first six months or so after we built it at the beginning of last year.
现在，您可能会想知道我们如何知道它是有效的。我们可以通过观察鸟脑是否进行准确的预测来确定这一点。现在，这张图显示了我们去年年初构建鸟脑后前六个月左右的预测质量。
This particular chart is specifically for the course teaching English to Spanish speakers, but we got almost identical results for all of our courses. Now, it's important to note that not only is the prediction quality high, but it's increasing and improving over time as more learners use the platform.
这张特定的图表是为教授英语给西班牙语母语者的课程制作的，但我们得到的几乎相同的结果适用于我们的所有课程。现在，重要的是要注意，预测质量不仅很高，而且随着越来越多的学习者使用该平台而增加和改善。
Another thing that we do, and we do this for almost all of our product changes, is AB testing. Now, the way this works is we take all of our learners and split them up randomly into two different groups.
我们所做的另一件事，几乎针对我们所有的产品改变，就是AB测试。这种测试的方式是将我们所有的学习者随机分成两个不同的组别。其作用是测试新产品的效果。
Now, in the case of bird brain, one half of those learners get lessons that are constructed using pre-existing heuristics that we've developed over the course of years. This is the control group, and for the other half of our learners, they get custom personalized lessons using bird brain. This is the treatment group.
现在，对于“鸟脑”的情况，其中一半学习者接受的课程是使用我们多年来开发的现有启发式方法构建的。这是控制组。另外一半学习者接受使用“鸟脑”定制个性化课程的治疗组。
And what we can do is compare learning outcomes between these two groups and see if bird brain is having a significant impact. So here are the results from one such AB test where we used bird brain to personalize level one lessons. And here we saw a 3.5% increase in content length.
我们可以做的是比较这两个群体的学习成果，看看“鸟脑”是否有显著影响。以下是一项AB测试的结果，我们使用“鸟脑”来个性化第一等级的课程。在这里，我们发现课程内容长度增加了3.5%。
So what this means is the sentences that you needed to translate or transcribe as part of the lessons were on average 3.5% longer. And this is one of the many different measures that we look at of how difficult the lessons are. Now, not only were they more challenging, but also more motivating because we saw a 6.3% increase in the time spent learning.
这句话的意思是，你需要翻译或记录的课程句子平均长度增加了3.5%。这是我们衡量课程难度的众多不同指标之一。现在，不仅更具挑战性，而且更有动力，因为我们看到了学习时间增加了6.3%。
So these learners were spending more time in the app. And we saw these same results again when we used bird brain to try and improve skill practice. We saw an almost 9% increase in content length and a 3% increase in time spent learning. And this is a pattern that we saw over and over again.
因此，这些学习者在应用程序中花费更多的时间。当我们使用Bird Brain来改善技能练习时，我们再次看到了相同的结果。我们看到训练内容长度增加了近9％，学习时间增加了3％。这是我们一遍又一遍看到的一个模式。
In this chart, each dot represents a different AB test that we ran trying to use bird brain to improve the learning experience in some way. And as you can see, in most cases, we were able to increase both content length and time spent learning. And in fact, there is a positive, statistically significant relationship between these two, meaning that a bird brain succeeds in making the lessons more challenging. It's also likely to succeed in making it more motivating by a proportional amount.
在这张图表中，每个点代表我们运行的不同AB测试，试图通过使用鸟类大脑在某些方面改善学习体验。正如您所看到的，大多数情况下，我们能够增加内容长度并增加学习时间。事实上，这两者之间存在正相关的统计显著关系，意味着鸟类大脑在使课程更具挑战性方面取得成功。同时，它还有可能以相同的比例使它更具有动力。
Now, this may seem obvious or even easy, but it turns out it's really not. So let's compare these green dots to these blue dots that represent AB tests that were run using more conventional product development and software engineering techniques. So for these other non-bird brain AB tests, there's almost no relationship between content length and time spent learning and a lot fewer of those experiments succeeded in improving both.
这可能看起来很明显或者很简单，但其实并不是这样。让我们把这些绿点和用更传统的产品开发和软件工程技术运行的蓝点进行比较。这些非鸟脑AB测试中，内容长度和学习所花时间之间几乎没有任何关系，而成功同时提高这两个方面的实验也要少得多。
But bird brain, because it takes into account both the difficulty of the exercises and where you're at on your own personal learning journey and puts the two together, it's able to strike a good balance and it balance the tensions between these two extremes and improve the language learning experience.
但是Bird Brain之所以能够在语言学习中发挥作用，是因为它考虑到了练习的难度以及你个人的学习水平，并将两者结合起来。这使得它能够找到一个良好的平衡，并平衡这两个极端之间的紧张关系，从而改善语言学习体验。
Now, today bird brain is used in some way for almost all session types that are in the app. And it's used in particular to choose the really difficult exercises that go into hard mode practice for extra XP or for the legendary skills that you'll also be hearing about today.
现在，几乎所有的应用程序会话类型中都使用了“鸟脑”，并且特别用于选择进入难模式练习的极难练习，以获得额外的经验值或传说技能，今天您也将听到这些技能的介绍。
And of course, we're working to make bird brain even better. We have ongoing research to make it more accurate, more linguistically detailed, so not only is there a 90% chance that you'll get this right, but if you get it wrong, it's probably because of this tricky noun agitive gender agreement, for example. And all of this makes bird brain more useful.
当然，我们正在努力改进"鸟脑"。我们正在进行持续的研究，使其更准确、更详细地进行语言处理，这样不仅你正确的概率达到了90%，而且如果你错了，很可能是因为如名词或形容词性别的复杂问题。所有这些都使"鸟脑"更加有用。
So if you're excited or interested to learn more about the ways that we're using artificial intelligence, machine learning, natural language processing, and so forth here at Duolingo, I encourage you to check out Duolingo.ai.
因此，如果你对我们在Duolingo使用人工智能、机器学习、自然语言处理等方面的方式感到兴奋或感兴趣，我鼓励你访问Duolingo.ai，了解更多相关信息。
Thank you very much and enjoy the rest of Duolcan.
非常感谢您，愿您享受接下来的Duolcan。