Blog Post

How Mia Learns Kids’ Interests

September 21, 2018

How Mia Learns Kids’ Interests

Conversational AI through Natural Language Understanding, Machine Learning, and Open Learner Models

By Darren Cambridge

When making book recommendations, Mia looks carefully at each child’s interests. There are good reasons for that: Reading comprehension increases when young people read books that interest them, and having a clear understanding of their interests increases their psychological well-being and sense of purpose.

But how does Mia know what a child’s interests are? Most reading software, if it considers interests at all, simply asks readers to select broad topics or genres from pre-determined lists. At Mia Learning, we treat interests as both broader and more nuanced. They are broader in the sense that they include not only interest general themes or topics of books, but also characteristics such as the type of characters they contain or the writing style used by the author. They are more nuanced in the sense that they include finer-grained topics.

Mia can’t ask a child to pick from a list. That would be a crazy-boring, hour-long conversation, at least. Capturing interests effectively requires a more sophisticated approach. Mia uses two sources of information: what kids say and what they do. Just as a teacher, librarian, or parent would, she listens and observes. However, since she relies on artificial intelligence (AI), these two activities take different forms.

Understanding what kids say about their interests

The first conversation kids have with Secret Agent Mia is the Briefing. Mia asks the children about themselves as readers, building a casefile for each child that she will then use to full her mission of finding great books on their behalf. One of the first questions Mia asks is what interests them: What do they love to read, read about, or learn about? A child might say something like:

“I’m really interested in horses, and Quasimodo amphipod, because Ms. Hare told me about them, and also, I usually love books that make me and my friends laugh.”

To understand this statement, Mia needs to parse and identify key phrases that are interest clues, such as “horses” and “make me and my friends laugh.” She then needs to determine which of the large set of topics, genres, and book characteristics she knows about they express. Here, “horses” suggests the child may be interested in books about “animals.” It may also suggests the child might be interested in “animal fantasies,” a genre in which horses are frequently featured. “Make me and my friends laugh” may suggest a preference for books with a “funny tone” and for “humorous stories” that are focused on making the reader laugh.

These terms are part of the set that professional librarians used to classify books in Mia’s database. They arrange the terms in a hierarchy, so that Mia knows “horses” are a type of “animal.” To identify these key phrases and make the associations, Mia uses what natural language understanding (NLU) engineers call Named Entity Recognition (NER). NER is a process through which software determines that phrases within a text indicate a reference to certain type of thing, or entity. For example, the phrase “horses” refers to the entity “animals.” The simplest way to identify a named entity is to search for verbatim matches between a predetermined set of words or phrases that are reference terms for the categories, or words or phrases that mean basically the same thing, called either paraphrases or synonyms. Mia tries this first.

Because horses are a common type of animal, frequently featured in children’s books, “horse,” is a reference term within the Animals entity, and a variety of ways of saying it—e.g., “animals you ride,” “ponies,” and “stallions”—are paraphrases. The system also automatically takes into account differences in number (“horses” is equivalent to “horse”) and verb tense (“made” is equivalent to “make”). Mia Learning worked with a former Google engineer to develop software to crowdsource the list of paraphrases kids might use, which our content experts refined and our engineers enhanced through adding semantically-similar words from general purpose lexical databases.

This approach, which computer scientists have been using for decades, is limited. For example, the Quasimodo amphipod is a type of insect discovered just this year. Given that it was coined so recently, it’s not surprising that the phrase “Quasimodo amphipod” is not a reference term or synonym listed within the “animals” entity. Generating and maintaining an exhaustive list of every possible animal and every possible way to refer to it would be a Sisyphean task. Mia needs to have a sense of what the names of animals look like in general, not just to know many such names.

The title of the broader, but still quite obscure category of “amphipod” might be a term on Mia’s list, but she also needs to be able to understand that in this instance, “Quasimodo” refers to a specific type of amphipod rather than the hunchback character from Hugo’s famous novel. Similarly, she needs to know “Hare” refers to a person, not a rabbit. In other words, Mia needs to be able to recognize phrases as instances of entities based on the context of the child’s larger statement about interests.

To enable Mia to identify unusual interests, taking context into account, Mia Learning uses machine learning. We have developed neural networks that extend Mia’s ability to do named entity recognition. (More specifically, Mia uses deep convolutional neural networks with residual embedding.) We trained these networks, which build on a general statistical model of English usage, using a large set of real world texts in which people talk about books. Our team annotated all the phrases within these texts that correspond to the types of interests for which Mia listens. Our machine learning technology used some of the annotated documents to build a statistical model and tested its ability to correctly identify the named entities in the remaining ones. It then went through thousands of iterations of adjusting and testing the model to maximize its accuracy.

Determining Interests from What Kids Do

In addition to listening to what kids say about their interests, Mia observes what they do. Anytime a child does something using the app that may be relevant to understanding the child as a reader, Mia records it as an experience (using the Experience API, also known as xAPI, format). Many of these experiences provide clues about children’s interests. For example, suppose the child Mia knows is interested in animals and humorous stories from the statement we have been dissecting chooses C.S. Lewis’s novel The Horse and His Boy from Mia’s recommendations. This action suggests that the child may be interested in books with similar qualities to those of The Horse and His Boy. Although the child may not yet be able to say so, they are likely to be interested in the genre “fantasy fiction” and in books with “courageous” characters. While reflecting on their reading with Mia, if the child reports a highly satisfactory experience with the book, this is an even stronger indicator of these interests.

Refining Interests Through Reflection

Through the process of observation of relevant experiences, Mia learns more and more about children’s interests as they choose books, read them, and reflect on the experience. However, Mia needs each child’s help to ensure that her profile of their interests is accurate. In some cases, Mia’s hypotheses about a child’s interests based on their experience records may be off base. For example, the horse-loving child could have just liked this particular work of fantasy fiction because of its themes or characters and not have much interest in other books from that genre. A child’s interests are also likely to change over time. The early preference for humorous stories might fade as the child increasingly chooses to read for purposes other than entertainment, such as to learn skills related to leathercraft, a new hobby.

To ensure that her understanding of interests is accurate and up-to-date, periodically Mia discusses what she thinks she knows about the a child’s interests with the child. Children can correct Mia’s mistaken conjectures, disavow stated interests that have faded over time, and share new ones. Researchers call using this approach an open learner model.

Through ongoing observation and discussion, Mia develops an increasingly sophisticated picture of a child’s interests that enables her to make increasingly effective book recommendations and to provide increasingly personalized coaching. The child also benefits from this process directly as their understanding of their own interests sharpens through reflection and their interests grow broader and deeper through exposure to new books. I hope this post has helped you understand how Mia uses cutting edge AI technologies to make this possible.

Uncategorized

How Mia Helps Kids Grow As Readers With AI Through Being An Open Book

June 26, 2018

How Mia Helps Kids Grow As Readers With AI Through Being An Open Book

Or, why kids actually do have to understand artificial intelligence to use Mia

By Darren Cambridge

Mia is designed to be so easy that even a six-year-old can use her with little or no adult assistance to support their growth as a reader. Children who are only beginning to read can happily have long conversations with Mia and have used her to select and reflect on books they love. So, this may come as a surprise: Kids need to understand artificial intelligence to use Mia. 

While there are some intriguing experiments underway to teach artificial intelligence concepts to younger children, AI has generally been the subject of college level computer science courses. So how can we possibly assume that an elementary-school student could understand it?

Thankfully, kids don’t need the same kind of understanding that engineers do to fully benefit from Mia. They don’t need to learn the intricacies of supervised learning algorithms or natural language understanding libraries. They don’t need to know how to control servos or temperature sensors (although all of us here are robotics club fans).

Kids do need to understand:

  • That Mia is not human. She has a number of impressive strengths and significant limitations, and she’s never going to be a replacement for the knowledgeable and caring adults who contribute to kids’ lives as readers.
  • What Mia believes she knows about them as readers and the reasoning behind her recommendations and suggestions.

Mia As Droid

Researchers such as sociologist Sherry Turkle have suggested that artificial-intelligence-powered toys and personal assistants have the potential to harm children’s social development, noting that some children don’t seem to make a distinction between their relationships with such system and real human relationships. Like the developers of the technology that Turkle examined, we do hope that children develop an effective bond with Mia, developing trust in her helpfulness and honesty and enjoying her corny jokes.

Unlike them, we never want kids to forget that Mia is a computer program or to think her presence in their lives eliminates a need for teachers, librarians, parents, or (human) friends. Mia is explicit about what she can do well—she knows more about almost every book in print than is probably possible for any human to remember, and she’s available 24/7 and for as long as they need her—but also about her limitations. For example, when Mia can’t understand something a child has said to her, one of the ways she replies is:

OK, so here’s the thing. Some things are easy for me that are hard from you, like remembering how many pages there are in each of thousands of books. But there are other things that are a breeze for you and really tough for me, like understanding whatever it was you just said. I’m going to get better everyday, but it will be a long time before I’m even as good as a kindergartner. For now, think about it as a game: What ways can you say what you want me to know or to do that I’m actually smart enough to understand?

Along the same lines, Mia is intentionally cartoonish in appearance, not photorealistic. Her voice is expressive, but isn’t going to fool anyone on the phone. We aren’t creating a virtual reality. Mia isn’t trying to pass as human. Mia is an explicitly virtual presence firmly situated within, and in service of, kids’ real worlds. She draws attention to her own artificiality.

The best analogy for what we’re aiming at is the relationships between people and droids in Star Wars. Droids have enough personality and contribute sufficiently to achieving goals that humans get quite attached. At the same time, no one confuses a droid with a person. Their limitations are self-evident—R2D2 can’t speak a human language—and their appearance is distinctly mechanical—human-shaped C3PO is gold and clunky. They have non-human talents, such as talking directly with space station computers, speaking three million languages, and projecting images of distressed princesses…but we digress.

Droids are an integral part of the central characters’ world, but they collaborate with and in service of humans (and other sentient beings), none of whom see droids as a substitute for other living, breathing people. Even as the technology that powers her grows exponentially more powerful over the coming years, we always intend for Mia to be more like a droid from Star Wars than a replicant from Blade Runner or a host on Westworld.

Mia As Open Book

Mia is dedicated to working her way out of a job. Mia’s goal is not to make children dependent on her to make decisions for them. Rather, she seeks to directly support their decisions today while also helping them learn to make such decisions independently tomorrow. Children are more likely to choose books that work for them with Mia’s help, but as a result of working with her over time, they should also learn to make better choices when she’s not around.

One of the ways Mia supports this growing independence is through scaffolding. In education, the term “scaffolding” means providing support that simplifies a task sufficiently for a novice while also modeling how experts tackle it. Scaffolding provides insight into questions like: What kinds of information do experts examine? How do they use it to predict how things will turn out? How do they makes sense of what they have experienced?

As the novice becomes more skilled, scaffolding can be gradually removed, much like literal scaffolding is taken down as parts of a new building are completed. In other words, effective scaffolding not only makes a task easier, but also helps someone learning it understand how to do it well on their own.

Through sophisticated AI, Mia finds a few books to recommend to a child from tens of thousands she knows about, drawing on what she believes she knows about the child and her expert-informed understanding of how to choose books. Unlike other reading software, Mia doesn’t stop there. She not only simplifies the task—here are six books to consider, rather than a ten thousand—but she also performs scaffolding through modeling. She explains the reasons she believes a recommended book is a good fit for a given child.

For example, here’s how Mia might explain the reasons why she thinks a middle school girl will love Pharoah’s Daughter, by Julius Lester:

  • You’ve said you’re super interested in Egyptian mythology and have read several books about it; you’ll enjoy the challenge of reading this book.
  • You often like books with detailed illustrations and stories about family relationships, and this book has both of those elements.
  • You might enjoy reading a novel featuring mythological characters to complement the non-fiction you’ve read so far.

Here, Mia is applying several beliefs often applied by expert readers:

  • If you’re really interested in the topic of a book and have related background knowledge, then you’ll be able to handle a book that’s more difficult than what you typically read and will enjoy the challenge.
  • Your preferences matter: You’re more likely to enjoy books that employ formats, plot structures, writing styles, and so on similar to those used in books you already like.
  • At the same time, variety increases satisfaction. Expert readers seek out books not only with similar topics and characteristics, but also with differences that broaden their literary experience, including through reading multiple genres.

Mia also is being transparent about what she believes she knows about the child as a reader that informs her judgment. Mia think that she understands something about the child’s interests (Egyptian mythology), preferred book characteristics (detailed illustrations, family stories), and experiences with genre (have read non-fiction). Mia also shares the sources of information that inform these inferences (you told me; you’ve read and liked books like this).

Educational technology researchers call this an open learner model. In AI software that uses an open learner model, the system actively shares what it believes it knows at a given point in time about the learner and what it has observed that informs those conclusions. Learners have the opportunity to reflect on how the system’s beliefs about them compare with their own. They may sometimes also have a chance to challenge the systems’ assertions, providing additional information that can lead to changes in those beliefs. For example, if a student believes they have a better understanding of a mathematical concept than the system believes they do, the student could successfully answer a few quiz questions to prove their point.

Mia also embraces this second dimension of the open learner model, albeit with the child themselves—not yet another test—as the authority. When students debrief with Mia after their mission—that is, talk with her about how well the books she helped them choose and how their time reading them has worked out—they share additional information with Mia about their reading experiences, beliefs and preferences that may differ from what Mia thought she knew about them. Mia invites them to reflect on targeted aspects of their experiences that may enrich both her and their understanding of how to make future choices. Mia learns from these conversations, improving the quality her subsequent recommendations and coaching.

Mia truly is an open book, constantly being revised in dialogue with the children who use her. Like a book, children consult her with a clear understanding that she was crafted by people and is limited in scope and function. But like a book, and through books, she has the potential to open up new possibilities for imagination, joy, and discovery.