Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Google Podcasts | Stitcher | TuneIn | RSS
Aaron BradleyAaron Bradley uses knowledge graphs to create intelligent content experiences at Electronic Arts.
Knowledge graphs are a relatively new technology that lets content engineers create more meaningful and versatile content experiences.
We talked about:
- his work as a knowledge graph expert and content modeler at Electronic Arts (EA)
- his definition of a knowledge graph: “machine-readable facts about things derived from one or more sources”
- the role of knowledge graphs in content work
- the two main benefits of content graphs for intelligent content: providing structure and assigning meaning
- the need for different ways to handle content to execute omnichannel strategy
- the extra work involved in precisely describing content but also the benefits of that work
- the use of knowledge graphs to facilitate personalization
- the importance of being able to connect your content
- how knowledge graphs can help connect content across organizational silos
- the makeup of the team he works with at EA
- the meaning behind his title “knowledge graph strategist”
- how to prepare to become a content knowledge graph specialist
- the arrival of headless CMSs as a content management option
- the not-immediately-apparent ubiquity of knowledge-graph technology already
Aaron’s bio
Aaron is a knowledge graph strategist at Electronic Arts. He is a linked-data enthusiast, search veteran, and compulsive categorizer.
Connect with Aaron on social media
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Content Strategy Insights podcast, episode number 104. The content industry is in the midst of an exciting transition, from old-school publishing and media workflows to new practices that create, manage, and present content in dynamic and more efficient ways. This new world needs content that is thoughtfully structured and meaningful to both the humans who use it and the machines that manage it. Aaron Bradley and his colleagues at Elecronic Arts use knowledge graphs to craft this kind of modern, intelligent content.
Interview transcript
Larry:
Hi, everyone. Welcome to episode number 104 of the Content Strategy Insights Podcast. I’m really happy today to have with us, Aaron Bradley. Aaron is a knowledge graph strategist at Electronic Arts, the big game company. And welcome, Aaron. Tell the folks a little bit more about your work there at Electronic Arts.
Aaron:
Hi, Larry, a pleasure to be with you.
As my title suggests, I am indeed someone who deals with the strategy and then application of knowledge graph technology. The less glamorous part of my job, and something I’m sure that we’ll get into in some depth, is as a content modeler. So I take conceptual models developed by our ontologist and convert them into logical models for use in our content ecosystem. So dealing a lot with modeling, but over the past many years I’ve been involved with what was at one time called semantic web technology morphed into linked data, and now is typified by the phrase in all of its permutation, knowledge graph.
Larry:
I think, I don’t know how common the knowledge is of the knowledge graph concept. Can you define that for folks what a knowledge graph is?
Aaron:
Sure. There’s about two dozen definitions of knowledge graph, lingering around there. So I’ve developed one of my own, what I hope is somewhat comprehensible, and a knowledge graph is a knowledge base of machine readable facts about things derived from intelligently connected, facts about things derived from one or more sources. Each of those phrases in that is doing a lot of work, especially the intelligently connected. So knowledge bases are something people are well familiar with, and you can even think of some sort of relational database tables or spreadsheets as being a type of knowledge base.
But I think the big innovation of the knowledge graph is that facts in different repositories or even within the same repository are intelligently connected. That is to say the relationships between things that is concepts and real life, things, are precisely defined so that machines can understand it.
Aaron:
So you might say, “Aaron’s father is David.” And it’s actually that relation is very precisely defined. So in the non-knowledge graph world, you might have a table that has Aaron and David there as entries in a cell, but it’s really, “Aaron has father David”, and then, “David has son Aaron”, and then, “David has wife Shirley.” And then you can start to connect all of these different facts by dint of those relations and start to do really interesting things. So as many have opined before, I think the main innovation of knowledge graph technology is that those relationships become first-class citizens. How things are connected are as important as those things themselves.
Larry:
Right, because in knowledge graph construction, there’s as much meaning ascribed to the connections as there is to the entities that are connected.
Aaron:
Exactly, so.
Larry:
And that’s part of the… Well, now let’s connect that to content, because this is content strategy podcast, and I know I’ve seen you present a few times on your work at EA around… Like the game rating example, but I’m sure there’s others as well. In how many ways, are you able to apply your knowledge graph work to the content at Electronic Arts?
Aaron:
I think in multiple ways, but I think it all hinges on really two aspects of what these days we probably call intelligent content. And that is what these technologies, the benefit that these accord content is that you’re providing that content with structure and meaning. And those are always the two main points of whatever it is you’re enhancing, that content ecosystem with in any enterprise, or even very small concerns, is that the difference between kind of traditional content management and what one could call a content graph or semantic content, is that your content has structure so you know the elements it’s composed of, how those elements are connected, and how they can be used together.
Aaron:
And it has meaning, and that’s the semantics of content. So it has meanings both by having those elements to find ,something like I have line or an image. These have, as a result of that structuring, they have some meaning. And then there’s also explicit declarative meaning such as the type that is provided by taxonomies, so that a piece of content is going to be about a particular video game. It’s going to be of a certain… got all of these different attributes of a piece of content can then be precisely described, and again, in a machine-readable fashion. So making it useful for humans by ensuring consistency for machines and embodying your content with structure and meaning in order to make it more usable.
Larry:
And one of the things that drives that need for additional uses is, well, the modern sort of content landscape. Let me back up a little bit, I think the most common way that people, when they think about it in the content strategy world, the most extremely structured content, I think, would be in the technical content strategy world with DITA, and XML, and the very standardized presentations.
Aaron:
Correct.
Larry:
But there’s a need for structured data beyond that. And, I think, to the grossest oversimplification of structured data would be like form fields in a CMS that looked just almost exactly like the presentation layer on the website. And so, I think, that maybe describes the continuum, but how do we get farther along that continuum to more where the structure is sort of abstracted, or the content is abstracted from its end use, I guess?
Aaron:
Sure. Well, and actually you raised an interesting and important point in terms of the demands on modern content strategy. And increasingly, and certainly for any enterprise of any size, that’s omnichannel strategy. So once upon a time, content was produced very much with the front end in mind, what content looked like was a part of that structure and part of that meaning. But as we moved beyond the web into the world of applications, into the world of API-driven content, that becomes far too restrictive and makes that content not as usable across as many different end points as it could be. So in this act of structuring, what you’re doing is separating the presentation layers from the data layers, and that allows you to have flexibility with both of those, and to be able to present content for any number of different end points, whether that consumption environment is a webpage, an application, a mobile application, a smartwatch, a speaker. It doesn’t matter once you’ve separated the data from the content because the data can flow through to that end point.
Aaron:
And that has really driven a lot of the change in content management in particular the innovation of the headless CMS, which is kind of basically what that means that you’re formatting information in terms of visual display, isn’t provided along with the data itself, that those concerns are separated.
Larry:
You reminded me of a phrase you used at OmnichannelX in your presentation, you said channel agnostic content.
Aaron:
Exactly, so.
Larry:
I love that, that’s a great way to describe that.
Aaron:
And when you look at the semantics of something, let’s say an image is about the game Battlefield V, that’s true regardless of where that image is used. It doesn’t matter whether it’s used in a social media post, within a webpage, described on a speaker, the semantics remain the same. And once you’ve described that, you get a lot of mileage out of it. And I’m pointing this out to say too, that it may seem burdensome to describe content with a really high degree of precision and using things like controlled vocabularies and taxonomies. But once you do that, you’re able to be a lot more flexible with that content, how it appears and how it’s combined with other pieces of content. So something like an image, which is a content element in a typical kind of piece of content, like a webpage, once you’ve describe that and structured it, it gives you a lot of flexibility as to where and how you use that.
Aaron:
And the act of, say, classifying that in terms of what it’s about just as a fundamental classification means that in other processes you can say, “Get me images about Battlefield V.” And you can create new pieces of content and new content products based on what you already know about your content. Whereas, if that’s bundled in just a content blog where the only way you know what that image is about is because of the context in that particular piece of content it appears, then the benefit of that image is going to be limited to the piece that it is used in, and it essentially becomes not very easily reusable.
Larry:
Yeah. That’s like, well… I’m thinking again the OmnichannelX when Noz Urbina talks about this kind of content, he talks about a Lego kit instead of a sculpture, and you’re talking about the Lego. So if you have a kit that has a bunch of different pieces in it, and they are connected in certain ways, they have a method by which they connect, like you have both structure and meaning, you can do all kinds of stuff.
Aaron:
That’s a really good analogy.
Larry:
I stole it from Noz. Yeah. Well, hey, I think one of the classic cases, and it’s also kind of a modern buzzword, is the notion of personalization, like how can you help give each individual ideally their own personal experience on the web? But tell me how you can… It’s kind of easy to see how this could help with that, but specifically what are some implementations of the use of knowledge graphs to facilitate personalization?
Aaron:
Another great topic, and you’re definitely onto something here. Personalization has long been a buzzword within the digital enterprise and has been every marketers dream for some time, because the more personalized content is the more in which it is tailored to the specific individual consuming that, or possibly consuming that content, the better the engagement is going to be. And over the course of time, in mercantile enterprises in any case, the more money that you make, so there’s lots of strong incentive to figure out personalization.
Aaron:
But the how has, until recently, often been fairly elusive and very manual. And it all boils down to, you need to know two things in order to personalize content. You need to know what your content is about and you need to know things about the user that could be consuming that content. And that’s where this suite of technology starts to bridge that gap, because you can use a common language to bind those two things, so that if you know…
Aaron:
Well, I’ll just keep with the Battlefield V example, I’ve got a bunch of content, what content might appeal to a certain player. And I need to know something about that player. And let’s say the one thing I know about that player is that they enjoy first-person shooter games. Like classifying a piece of content about Battlefield V without having to tag that first-person shooter game, because I’ve classified that as Battlefield V, and Battlefield V is video game, has a genre, and one of those genres is first-person shooter, that gives me the ability to speak the same language as the player that may be interested in that sort of content.
Aaron:
So in determining at the machine level, what sort of content do I serve up to someone that’s interested in first-person shooter games, it’s going to be something about a first person shooter. And that semantics in this case gives you the ability to affect that personalization. It allows you to bind the interest of the user, or desires of a user, or needs of a user with the content that you have available and provide that to them. So it’s an actual mechanism to facilitate personalization rather than, “Wouldn’t it be great if we could offer more personal content, how are we going to do that?” And there is a way.
Larry:
That gets into the meat of personalization, like actually connecting that. And so that implies a lot of cross disciplinary work with marketing people who have CRMs, or wherever that customer information is coming from. And then maybe a little more work, but worthwhile work on the content side to ascribe meaning and intent and purpose to the content that you’re creating. How did those practices evolve at EA or other places that you’re aware of? How do you stitch that together?
Aaron:
Yeah, it’s always interesting points here, Larry. And that is really at the heart of both opportunity and challenge within the enterprise, because typically these different units, the marketing organization, those that produce media objects, those that produce content, those concerned with merchandising, operate in fairly independent silos, and they’re accustomed to operating in those silos. And that’s been, to a certain extent, fine because there’s been no robust way of starting to connect those different silos within the enterprise. But once you start to develop a common language in the form of those conceptual models and those knowledge bases that combine data from different sources that now able to do so, because they share that knowledge, it gives you the ability to break down those silos.
Aaron:
The challenge is that, institutionally, that can be very difficult because indeed these different units within any big enterprise may not know of each other’s existence, let alone how that interoperability is now being facilitated.
Larry:
One of the things that led to that question is, in your talk with OmnichannelX, you mentioned that EA has an ontologist on staff, somebody who-
Aaron:
There’s plural.
Larry:
Oh, plural, more than one. Okay.
Aaron:
And we now have an ontology staff.
Larry:
Oh, sweet. Okay. Well, that’s even more interesting then, so there’s a team of people concerned. Is there… Like what I imagine in my dream world is an enterprise-wide alignment on language, and that these ontologists were working some kind of magic to do that. How close are we to my dream of that happening, I guess?
Aaron:
Yeah. It’s not that far, but I maybe have a different take on that. And that also pertains to what I was talking about with silos is that… And here I’ll crib a line from Kendall Clark of Stardog, the knowledge graph platform provider, “Unconnected data is a liability.” And by facilitating the connection between data using… And there’s all sorts of knowledge graphs, basically, but that conceptual layer of common understanding at some form or another, it’s not that your forcing people all to speak the same language. “Here’s Esperanto, everyone speak Esperanto and English will now be prohibited.” What knowledge graphs provide, or subsets like content graphs, is a mechanism by which diverse sources can be treated as if they were one. And you often hear the term data virtualization in this regard.
Aaron:
So whether or not you’re natively building processes and applications against the common language, let’s say that you’re developing. Probably more important than that is the ability to tap into resources that are already there, typically in relational databases, say, “Oh, cool. Now that I have this language that describes a car…” You may have described in your table an automobile a different way. You may say it’s a wheel and this other language may say it’s a tire, and another language may say it’s a round rubber device. But once you know that you’re all talking about the same thing, you can take data from three, or four, or five, or 10 different sources about automobiles and treat them as if they were all the same, and do interesting and useful things with that data, because it is now connected. Whereas, before, it was unconnected and living in those silos.
Aaron:
So I just want to kind of underline that, and which is my definition says, that a knowledge graph is knowledge base, where the data comes from one or more sources. To me, the kind of big value proposition of knowledge graphs is the ability to combine data from different sources and be able to use it meaningfully. I think you saw that because you saw my presentation, I’ll tell your listeners, so that means that something like game ratings information, which was out there in a silo and unconnected to other processes, it’s like, “Okay, we need to be able to say that this game is rated M for Mature in this particular jurisdiction.
Aaron:
But that’s also data that once you have a method of describing that with a common language, you can say, “Oh, you know what? I can then match that”, and where this gets back to personalization, what if the concept of M for Mature… I forget the exact age range, which is for 17 up… Guess what also has an age? Our players have an age. And if you can say the age that in our player databases or marketing databases that says, “This is age.” It’s also, “Oh, this is age represented by a rating, and therefore, I can start to understand our users interest in regard to that data.” Whereas, before, it simply was used to produce the little icon that you see in a video game cover, it’s quite a lot different.
Aaron:
So the more in which you are able to provide that level of semantics to understand the meaning of your data in all of these different places, the more in what you can use and combine it. And people are fond of saying, “Derive new information that wasn’t possible before.” Let me give you a really quick example of that, that I think people can connect with, I may have one database that says… And I’ll use Celsius since I’m a Canadian… that the normal temperature for a human is 39 degrees celsius. And you may have another piece of information, “Larry’s temperature is 40 degrees.” Those are just two separate pieces of information, but when you combine them, you can say, “Larry has an abnormally high temperature and should probably go see a doctor.” So just having either of those two pieces of information, doesn’t send Larry to a doctor. It says, “Oh, humans should be at 39, Larry’s at 40.” You combine those and you say, “Oh, this is then out of range and should see a doctor, or maybe just at 40 to take an aspirin.”
Larry:
You’re reminding me of… Because that’s, I think, what they call in the ontological world, inference. So, that’s one of the things you can do.
Aaron:
Exactly.
Larry:
But that makes me immediately think of one of the things we all struggle with, and it comes up in almost every episode of the podcast is like, what do we call each other? And how do we work with each other? And how do our titles align with the actual duties that we’re doing? So tell me, your role as a knowledge graph strategist, and you work with probably information architects. You’re mainly a modeler. I guess, maybe help me orient you to your practice and the closely adjacent practices that you’re working with at EA.
Aaron:
Sure. But let me begin by talking about the other members of my team, which then may help frame that. As you mentioned, we have ontologists, so we have a lead ontologist and other ontologists working under him. So in our content ecosystem with, as we’ve morphed it over the years, now new modeling begins with our ontology. So we start at the conceptual level. If there’s a new content requirement, we take that view of it first, rather than a logical view.
Aaron:
So if you were to need to describe a different aspect of your business through your content, then we would say, “Okay, what’s the nature of that?” And then from that, and this is where I come in, we say, okay, the conceptual model, the ontology has said that, “This is the way the logical model looks”. That is to say, here’s the different content elements and how they’re connected. And also here are different controlled vocabularies, so whether those are authority records or taxonomies that need to be created to support this, which leads to another part of my team, which is taxonomist. And again, plural, we have a lead taxonomist and other taxonomists working under her, or working under contract with us to build required taxonomies.
Aaron:
And then finally, in order to get this stuff to work together in a modern technology stack, we also have a knowledge engineer that is capable of dealing with all of those acronyms and standards that are baffling to many – RDF, OWL, SHACL – how do you actually get this all to work? So in aggregate, we’re kind of like an information architecture shop. But instead, we deal with information architects in other units, we’re really a knowledge design and management shop, our small team.
Larry:
Got it. And you said, you’re… Like if somebody made you identify yourself, you might call yourself a modeler. And I think that’s a role that’s… I had Jeff Eaton on a couple episodes back, and you’ve talked about like title case content modeling and lowercase. And lowercase content modeling is like aligning stuff with your content management system, kind of that implementation level.
Aaron:
Exactly.
Larry:
Whereas at the higher conceptual level, he talks about like “the friends you meet along the way.”” It’s like, that’s about where you align human beings and get squared away on the concepts. Is that how you articulate content modeling? Do you see it in those kinds of terms, or how would you define content modeling and how you do it?
Aaron:
Yeah, I think from my perspective and my intersections with it, it’s very much at the technical level, so I’m designing very precise, based on those conceptual models, very precise… Think of them as, for the listeners that might be having problems comprehending the abstractness of this thing, blueprints of how then to build a content product. So a content model for an article is going to have a headline, and body, and one or more images, may have embedded audio or embedded video, and it’s going to have an author, and all these different elements. And what a content model says is, “Cool, these are the different elements of an article. The usage of those elements is, “Do you require an author or not? Do you require a headline or not? How many of these things you can use?” “Oh, only one headline, but I can have multiple images.” This all allows you then to, in another technology stack, configure a content management system to understand that and make use of that in terms of the content produced, most often in conjunction with the design system.
Aaron:
So in terms of my content modeling role, that’s very technical. And content… But that’s a good question because there are indeed the upper- and lower-case versions of that. And content modeling or content models can also be thought of, and this is not at all the way I think of them, but this crops up occasionally as being models of how content satisfies content strategy, for example. There’s another realm in which content models are, how to monetize content. And still another realm if you are shocked, as you start to Google, or look on Twitter for this, there’s also people that model themselves as content for money, typically in various degrees of nudity. So a content model is a kind of loaded word and kind of many meanings.
Larry:
Oh, that’s too great because I think the term content strategy has been appropriated for so many different meanings and now content modeling has already been appropriated.
Aaron:
And I guess I still haven’t addressed why am I called a knowledge graph strategist, which is a title I admittedly gave myself and I’m grateful my boss went along with that. And this may be instructive for some of your listeners, it’s been a multi-year journey. My boss, who is currently my boss, and I began kind of working on this approach as side projects off the side of our desks that we collaborated on over the course of time, that began to develop momentum. And that we began to articulate use cases that we were able to design proof of concept to see success with those efforts. And especially, to an earlier point you made Larry, to provide a blueprint for future personalization and recommendation because this technology works extremely well for this, and gaining traction over the course of time. And also looking at the environment around us in terms of other enterprises, we’re able to get more executive buy-in and then more resources to build out the ecosystem that we’re currently working on now. So part of that strategic role has been, over the course of time, both evangelizing and working on the fundamental approaches to what we wanted to do. But it’s by no means been a linear or uninterrupted journey, there’s been lots of ups and downs, and false starts and many, many, many lessons learned along the way.
Larry:
Yeah. Well, hey, one of those lessons I’m really curious about now, because I think with the increasingly decoupled nature of modern content and that like what you just went through at EA the last few years, it’s probably going to be happening all over the place. I wonder if you have any advice for content strategist, or content designers, or content operations folks who want to migrate more this way? What are the skills sets that you would develop to kind of migrate this way?
Aaron:
Sure. And actually in some ways we just address that, and then I will underline, put an asterisk, and boldface on this, professionals that are conversant in this space, if you are getting into the realm of graphs, look at ontologists, knowledge engineers, taxonomists, and people that have done this before in an enterprise environment. There’s still no blueprint, “Oh, I want to have a knowledge graph. I mean, now I have a bunch of relational databases, how do I go about this?” There’s no single way of effecting that. And there is, to this day, no kind of set way of approaching that, but you’re going to be a lot better off with specialist talent than not.
Aaron:
And in terms of what would I do differently earlier on as to where we are now is that I would have hired some of those key positions earlier in terms of ontologists and knowledge engineer and those with familiarities with graph databases and those standards by which… Because we’re a semantic shop rather than, that is a semantic graph, rather than a property graph shop, that means familiarity with RDF and OWL, and RDFS, and SHACL. SHACL is super powerful. Even aa saying that, this also takes place, it’s not merely enterprises catching up to tried and true technology. The technology is changing all the time too. SHACL, it’s a shape language very much like OWL, but where the emphasis is on shape validation, that is, is your data compliant to the schemas that you produced, has turned out to be enormously useful. But it’s only a few years old, right. Tools now have just over the past couple of years come to really be able to utilize SHACL.
Aaron:
So at the same time that enterprises are, “Oh, I’m going to move into this sophisticated knowledge economy”. Even as they are doing that, then the tools continue to progress, and the standards continue to change and evolve. And part of the reason you’re talking to me is through seeing me at a headless CMS conference. It wasn’t that long ago that I was like, “Hmm, how do we manage our content structurally?” And looking at CMS providers, and it wasn’t that long ago, there weren’t any, or weren’t many at all. So the headless CMS is a relatively recent innovation as well.
Aaron:
Just a few years ago, you’d be, as an enterprise, more often looking at the Adobes and other big enterprise monolithic content management systems. And to your point about decoupling, that’s also an aspect as you move into the sort of knowledge technologies is that single platforms tend not to predominate. You may have very important pieces of the puzzle, particularly vocabulary management systems, or taxonomic management systems, or headless CMSs, but it’s not like the AEMs of the world where, “Oh, we have a solution that is going to solve all your problems and all things are here.” Not only is that less likely now, but there’s also disadvantages to that. And that by decoupling each of those pieces, you’ll also allow those pieces to be used elsewhere in the enterprise.
Larry:
The way you just described all of that and all the stuff involved in that, if you’re a lifelong learner, this is the place to be.
Aaron:
Yeah, for sure.
Larry:
There’s no end of things to keep up with. Yeah. Well, hey, thanks there, Aaron. We’re coming up… I think we’re a little overtime, but that’s fine. I love this conversation, but I do want to wrap up, is there anything last, anything that’s come up in the conversation that you want to elaborate on, or just anything that’s on your mind about knowledge graphs, or content graphs, or…
Aaron:
I guess the one thing that I’ll note is they’re more ubiquitous than you might think, and that if you look at the modern enterprise, even those that don’t have named knowledge graphs, often have that technology under the hood. Here’s an example, and just because I’m a compulsive curator, and I look for this sort of information, a lot of this is exposed by job postings. So does Netflix have a content graph? Why, yes it does. Or else, why would they be hiring for someone to run their content knowledge graph? So it may not be publicly exposed, but it’s turned out that this sort of approach for the data rich, modern enterprise solves a lot of problems. It’s not there because it’s new and cool. It’s like, “Oh, how do I do personalization?” Well, here’s an approach? What are alternative approaches? Hmm, they don’t seem to do as good as job as this thing, let’s try and pursue this.
Aaron:
So knowledge graphs haven’t evolved because they have the appeal of the new and cool. They’ve evolved because they actually do meet business needs and are able to satisfy. Very compelling use cases and that’s going to be the case. I’ll wrap up, I guess, by saying… And I think I’m citing Dan McCreary on this, that this sort of approach is going to become more and more accessible over the course of time. That, right now, the growing pains that I’ve described and the mishmash of different technologies and having to figure it out as you go, is going to be less and less arduous as the space matures. And I think it’s going to be more and more likely for small and mid-level players where there’s a benefit to doing so of adopting this sort of technology.
Larry:
That’s what excites me about it is it seems I watch… You see the tools developing and the procedure… Just how to do it, just getting that knowledge out there is so much better. That’s great. Well, thanks so much, Aaron. Oh, one very last thing, what’s the best way if folks want to follow you? Do you have a preferred place on social media, or how can folks stay in touch?
Aaron:
Sure. I’m a Twitter guy. So, you can probably find me by my name, but my handle is @aaranged, but that’s with two As, rather than two Rs, like Aaron.
Larry:
Like Aaron, okay. Great. Perfect.
Aaron:
Yeah.
Larry:
Well, thanks so much, and I really had fun with our conversation.
Aaron:
Well, thanks a lot, Larry. It was a pleasure.
Leave a Reply