Katariina Kari: Knowledge Graph and Ontology Practice at IKEA – Episode 163

Podcast: Play in new window | Download

Subscribe: Apple Podcasts | Spotify | Android | TuneIn | RSS

Knowledge graphs let people and computers work from the same body of facts to create uniquely informative and powerful experiences.

Katariina Kari and her colleagues at IKEA use ontologies and knowledge graphs to drive applications like recommendation systems and to streamline back-end processes like image recognition.

Katariina balances her engineering expertise with a deep appreciation for the humans who create and use AI applications like knowledge graphs.

We talked about:

• her work on IKEA’s knowledge graph, in particular how their ontology brings meaning to their data model
• how web standards like RDF and OWL undergird ontology and knowledge graph work
• the benefits of codifying domain knowledge in a way that permits it to be used in multiple applications, e.g., both IKEA’s recommendation system and an internal image-recognition app
• their system to capture subject matter expert insight in the knowledge graph with intuitive authoring tooling for the SMEs
• the three-tiered view they take of knowledge graph construction at IKEA:
  • the ontology layer, which includes hundreds of class and property definitions
  • the categories layer, which includes thousands of controlled-vocabulary terms
  • the product/instance layer, which includes tens of thousands of specific instances of products and media asssets
• her strong feelings about the “human in the loop” and the need for codified, human-validated facts when building AI applications, as well as thinking that is grounded in the humanities and social sciences
• the need to stay critical in your consumption of information about AI, especially the current hype around ChatGPT and other LLMs

Katariina’s bio

Katariina Kari is the lead ontologist at Inter IKEA Systems B.V. Katariina holds both a Master of Science and a Master of Music and is specialised in the semantic web. At IKEA, she models the IKEA Knowledge Graph; a common vocabulary to offer customers compelling interior design solutions in the digital space. In her free time, Katariina advises institutions, such as chamber music festivals, in digitalisation, guiding the humanities to the digital age.

Connect with Katariina online

• LinkedIn
• Instagram
• Twitter

Blog posts and videos

• IKEA’s Knowledge Graph and Why It Has Three Layers
• The Art of Ontology
• Building, and Communicating, a Knowledge Graph in Zalando (video)
• Building The Fashion Knowledge Graph (video)

Video

Here’s the video version of our conversation:

Podcast intro transcript

This is the Content Strategy Insights podcast, episode number 163. Large language models like ChatGPT are getting all the attention nowadays. But another AI technology – the knowledge graph – is helping enterprises apply knowledge and facts about their products and customers to deliver powerful content experiences. Katariina Kari and her colleagues at IKEA have developed a system the captures hundreds of years of subject matter expertise to drive customer recommendation systems and automate and streamline internal processes.

Interview transcript

Larry:
Hey everyone, welcome to episode number 163 of the Content Strategy Insights podcast. I’m really happy today to welcome to the show Katariina Kari. Katariina is the lead ontologist at Inter IKEA Systems. So welcome Katariina. Tell the folks a little bit more about what being an ontologist at IKEA entails.

Katariina:
Hi Larry. Thank you for having me here in your podcast. That’s a lot of episodes, I must say. So being a lead ontologist at Inter IKEA Systems means that I’m basically in charge for how we shape the knowledge graph. So we’ve committed to, and we’ve been doing that for the past two years, to build a knowledge graph for the purpose of IKEA. And to capture IKEA know-how about interior design and people’s homes and furniture for the home, which we are selling in our big blue boxes in the potato field. And so as lead ontologist, my work is very much about the shape of the knowledge graph, so how it is structured, what its data model is. And when I say data model for our knowledge graph, the real word for that is actually the ontology.

Katariina:
So ontology is that because data model is a little bit too simplistic and ontology is really relating itself all the way back to Aristotle to the philosophical exploration of the… It’s not the meaning but the real thing about things like the essence of things and how to describe those. So what the ontology brings to a data model, it can not only describe what the abstraction level is and how they relate to each other, but it also gives you the possibility of reasoning. And by describing one fact you can describe many facts more. So the oncologist needs to actually pay a lot of attention to description logic. So when they’re creating a class hierarchy to make sure that if you instantiate from the lowest level you can instantiate the same thing from the highest level. That’s a first check that you need to do and then to name the properties or the relationships between things also accordingly.

Katariina:
So there’s a lot there. And I mean all of this started I would say 1999, but already, I mean description logic started already in the sixties and the seventies. But the semantic web started 1999 and that’s where these standards, RDF, resource description framework and the resource description framework schema, and then later OWL were developed to exactly give away in which you can describe class hierarchies and relationships between… that would materialize between the instances of these classes. And that’s really the craft of an ontologist is that they can dynamically describe anything. And now what is this anything they need to describe? Well, anything is human knowledge. So ultimately what an ontologist does is it analyzes the structure of the human knowledge and tries… and it’s always trying because there’s never a perfect answer, but it tries to create the best ontology with which this can be explained to the computer.

Katariina:
And explaining to the computer is also not the correct terminology. Ultimately it’s about how a computer or an application can in a best way utilize this human knowledge and automate on it. So that’s why also the description logic side is so important because by saying one fact, and that’s when I talk about the lazy ontologist. The lazy ontologist needs to describe an ontology in such a way that you need to only express one fact, describe many facts more automatically from it. And that’s what really ultimately makes it really available for automatic applications. In an earlier podcast, I was describing how the ontologist, also a really good colleague for software engineers because now the software engineers do not have to talk to humans and domain experts, subject matter experts. The ontologist create a system create to capture the subject matter expert’s knowledge, and then the software engineer just writes their code for the knowledge graph and for the knowledge base.

Larry:
No, as you described that, I’m laughing a little bit because I love engineers. My last boss called me “the engineering whisperer” and I love hanging out with them. But they are notoriously not as well suited to the human discovery work that designers are. Anyhow, I’ll just leave that aside, but you were just describing kind of what I’ve always understood is the top level benefit of ontology especially as expressed in RDF, which is that resource subscription framework. The format that gives… it’s like the HTML of knowledge stuff, I guess is grossly oversimplify, but that benefit of having things described in a way that makes sense to human beings but the computers can do something with – that seems super powerful.

Katariina:
Yes, that is very powerful and that actually… I mean if you don’t have that, you most likely have something like that in your code. So your code will already be trying to put a bit of human knowledge or you might have it in a database, but it’s locked from a general computability. So if it’s in an application code, if it’s in the code, you would need to… it’s really hard for any other application to reuse that. So that’s also the additional benefit or I think one of the greatest benefits is you’re taking business logic and domain logic, domain knowledge and putting it into a common repository from where it can be used by one application and reused by many applications before. One example that, we recently did that I was talking about at KGC where we were last time in May, in New York. We built a common repository for interior designer knowledge on if you’re buying a sofa, you can match that with decorative cushions, so that’ll make it more comfortable.

Katariina:
We created that and that drives the recommendations on ikea.com for when people buy a sofa, they get bunch of decorative cushions and throws and laptop stands recommended to them and lint rollers for cleaning. And now a completely different application than e-commerce recommendation. And this was computer vision came to us and be like, “Can we use this because this is really good?” Because if we’re recognizing furniture for picture and we recognize a sofa, then we can reduce the probability space of the other stuff that could be in this picture. Because most likely you will then find the decorative cushions and the throws in this picture and this will make their model more accurate. So these are two very, very different applications across the company, but now they can use the same repository. And this is a very beneficial effect because it means you have to build the knowledge graph only once and it becomes this knowledge infrastructure in your company where applications start popping up like mushrooms after rain in a forest.

Larry:
No, the generic benefit of codifying and articulating your human knowledge in the organization and then who knows what you can do with it. And that’s one of the generic benefits that knowledge graph people talk about all the time is like, “You’ll discover stuff that you never knew, like some other use or some other thing.” But one of the things when you did that talk at the Knowledge Graph Conference, we were talking before your talk and I just assumed that you’d be doing some fancy vector embedding fancy math on top of, you said, “No, we’re just taking applying business rules to this and making those recommendations.” And this is something that’s common in the content strategy world is working with subject matter experts to get stuff out of their brains and share it with people. You just do that in kind of a different way. Can you talk a little bit about how you work with subject matter experts and for that exact kind of task?

Katariina:
Well, in IKEA we have this really big vision, and this is really from our business owner, Adam Keresztes. He has this vision that we need to democratize the knowledge graph so that many subject matter experts inside of IKEA can have their say in the knowledge graph. So we’ve put a lot of attention to tooling and especially visual authoring. So you know how the knowledge graph, I mean it is a graph, so it has nodes and edges, it has boxes and lines between the boxes. And now what we’ve added is a subject matter expert who is non-technical can put a box into an empty canvas and draw a line to another box and this way author and add their facts into the knowledge graph.

Katariina:
So that’s been a really big work that we’ve been doing exactly in the past year to build that capability and really having these people who… I mean one of our colleagues, she’s been with IKEA for decades, for 20 years or something, and what she did is she started as a visual merchandiser. And she has that knowledge and now she’s expressing these interior designer facts, which she knows very well because she’s worked with IKEA for such a long time. She’s expressing them with a tool which now benefits so many customers in ikea.com and she doesn’t have to know how to code. And she still has challenges of wrapping her head around how computers work and how you have to do things on a generate level and not on the specific level, but she’s still doing that. And I think that’s a success story

Larry:
As you say that I’m reminded one of the reasons I wanted to have you on the podcast, you did this brilliant blog post a while back about the three-tiered knowledge graph at IKEA. And I think that that might be a way to kind of contextualize what you were just saying. I think it is that you have, for folks who don’t know that in the knowledge graph world, there’s traditionally just the A box and the T-box terminology box, and I forget the A box, but-

Katariina:
Axiom box.

Larry:
Oh, the axiom box, and then Dave McComb added this third level and you applied that to IKEA. Anyhow, talk a little bit about how you organized your knowledge graph at IKEA.

Katariina:
Yes. I found Dave McComb’s three layer pyramid for… Ultimately what it’s saying is how much of things are in the knowledge graph. So it’s saying that the top layer, the definitions, the definition part, the ontology part, that’s the smallest amount inside the knowledge graph. So it’s like in the hundreds. So you have hundreds of classes and property definitions, and so that makes sense to do manually and it makes sense for one team to do that manually. That’s what I applied if IKEA was saying, “Where do we need manual work? Where do we need automated work? Where do we need distributed manual work?” Or, “What kind of quality assurance do we need to do?” If it’s very little things, then the quality assurance can be very manual versus when it gets bigger it needs to be automated. And then what do we use that for? So I was just analyzing it for the many things that you want to do in a company to make sure that you have a good knowledge graph and also how to explain how to build one.

Katariina:
So the top layer is the ontology and that’s in the hundreds. That’s class and property definitions. And ultimately the shackle definition is there as well, like how things would connect and how you would want them to connect. And then the second layer, it is already the T box. It is already the instances, but these are very special instances. These are the categories. This is the controlled vocabulary of the company. So if you think about IKEA, for every product we say what its material is and what its color is, those are two very special things. And we only have tens of colors and not even tens of colors, 10 colors maybe, but we have a very, very small set of colors and also tens of material definitions. But then we have tens of thousands of products. So each of those products will refer to one color or two colors and one or two materials.

Katariina:
That means that there’s a lot of incoming links into one color, one particular color like yellow and a lot of incoming links to wood, for example. And so if you look at the graph from a graph structure point of view, these are nodes with a high page rank. So the second layer is really about those instances that have a huge page rank because everyone’s referring to them. And then if you look at a product, there’s not a lot of incoming links. There might be 10 incoming links because other products are connected to these products, but ultimately there are more outgoing links. They’re like a hundred category outgoing links to saying it is of this color, it’s designed by this person, it has this nice IKEA name attached to it. So many more like outgoing links that describe the product. So already from a node perspective and page rank perspective, I’m referring to Google’s page rank when they’re seeing that there’s a lot of incoming links, so they look very different.

Katariina:
So I call this second layer the categories layer, Dave McComb calls it the categories layer. I call it the layer of the specialized instances, the very special instances with a huge page rank. And that means ultimately it’s the terminology inside the company because somebody came up with what those materials are. I mean, we could be saying birge or pine, but instead we’re actually just saying wood, right? So somebody made that decision in the company and somebody stands behind it and somebody needs to change that if the company decides to be more descriptive about the wood material. And the same goes for colors. So color is a spectrum, but then we decide on certain terminology to a golden bowl is actually categorized as a yellow bowl. And so somebody made that decision and that’s why the middle layer is also the control vocabulary. And this needs to be in the hands of domain experts.

Katariina:
According to Dave McComb, and I agree with that, lies in the thousands. So you have thousands of terminology definitions inside your company so that you cannot manually manage with one team. The ontology was fine with one team, but the categories layer that needs to be distributed across the domain experts in the company. So the material people, they maintain the material taxonomy, so they are maintaining these in taxonomies and the color people are in charge of the color taxonomy, the style people of the style taxonomy and so on. That’s the second layer. And then ultimately you have the third layer, which is all the products, tens of thousands of products or all the marketing images we have, their metadata. Or all the content we create for the IKEA and its metadata in that connected form. So this content is about this product or it’s about this material or it’s about this color. Or this marketing image is showing the yellow clip and sofa together with this cushion and in this room setting and it’s a portrait format, it’s not landscape, et cetera. And that’s the connected data and that runs in the millions.

Larry:
Got it. And I want to run this by you, because I’ve contextualized this in my head as the top layer is the ontology, and that’s just concepts and stuff. And the second layer, I think a lot of people think about taxonomy as this driving thing across a lot of this kind of stuff. And I think of that categories layer is where most of the taxonomy happens. And then that bottom layer, the instance layer where just all the things you got and how those properties and the attributes of them and all that stuff is ascribed. And then it’s the stuff above that gives you the power to do stuff with the stuff at the bottom to say, “Oh, you looked at this couch, you’re probably going to love this rug.” That’s kind of how it works?

Katariina:
Yeah. So then what we do is when we say that in IKEA, when we say that, “Okay, every product can be categorized as… or some products are categorized as a sofa or more particularly indoor sofa because they fulfill these indoor criteria in terms of who designed it and the materials.” And then we have some products which can be categorized as a decorative cushion, and they are also suitable for indoors only specific ones of them. So that’s kind of described according to the category layer, like a business rule for how to find all the products that are indoor sofas and how to find all the products that are indoor decorative cushions.

Katariina:
And then those business rules, they are defined on the category level because those are your supercharged instances, but then you can have an application to materialize that on the bottom layer. And that’s the reasoning, the engine. And we are using Datalog for that, which is based on Prolog if you’re interested in technical application of it. But then according to those business rules, we are then materializing direct connections between each decorative cushion for indoors and each indoor sofa. And then they’re available for the common repository, be like, “Oh, these decorative cushion adds comfort to this indoor sofa.”

Larry:
Yeah, as you are talking about that, and I reflect on the last, I don’t know, 20 minutes, almost everything you talk about is there’s a person involved, like a subject matter expert or there’s people working on all this. And you talked something before we went on the air, you said something about there’s this thing in the AI and data and knowledge world about the human in the loop. But you were like, “We got to take the human out of the loop or properly put the human in the loop.” Can you talk a little bit about the people part of it?

Katariina:
Yeah. I feel very strongly about human in the loop because most of the time when we’re building… I mean in order to build artificial intelligence, we need to be careful about what are the things we can do manually and what are the things we need to automate. So in order for things to scale and to run as applications in a space, most of that needs to be automatic. And I think we need to be really careful when we are building artificial intelligence. And here at the beginning of building artificial intelligence, we need to be really careful about knowing where the human actually needs to be in the loop. So it’s not like were these automatic quality, like you said, quality assurance people are just checking afterwards if the computer did things correctly. But it’s actually that we need to have a vehicle to tell the computer a little little bit about us as humans, be like, “I know that the data suggests to you that this image of a dark-skinned person could be categorized as an image of an ape.”

Katariina:
That’s what your AI is telling you, but that’s racist, don’t do it. So we can actually give our ethics and our regulations and our law for the machine. We can also give it facts and be like, “You don’t really need to calculate a lot of data to come up with something. I can just tell you.” Like a cushion goes well with a sofa. And that’s not just about… It has even ecological and economical implications because it’s just faster and easier unless energy consuming to just say these things. But sometimes it can really makes sense to have a human in the loop because it’s just more energy efficient. And then there’s also a secondary practice to this. If humans are stating these facts, and by this I’m implying that they should store these facts in a knowledge graph like we do in IKEA. The secondary effect of that, it also makes the enterprise… It gives this practice inside the enterprise to really discuss and find the consensus of what do we mean when we say this?

Katariina:
Because there’s a lot of terminology definition that needs to be done. And to be said like, “What do we mean when we say a sofa is also a sofa bed or it’s made out of wood?” What we want to express and what do we mean when we say this? Or what do we mean when we say it is of a Nordic style or a modern style? What do we mean by this? And this conversation, which happens between humans, between professionals inside IKEA is very valuable. As a practice, it’s very valuable because it brings people together and lets them have this conversation and also lets them make decisions of this is where we want to go as a company. This is how we want to express ourselves. And now we can then put that into a knowledge graph and then build a better artificial intelligence with large language models or with computer vision on all these different things. But then now you have the good human in the loop in the middle of it, grounding these things in truth and not just a discovered truth, but actually decided truth.

Larry:
Yeah. As you’re saying that, you’re reminding me of two quotes just came to mind that I think will help stitch together the knowledge graph world and the content world. I just had a couple episodes back, I had Elizabeth McGuane on. She just wrote a book called Design by Definition, and one of the things she’s famously known for when she worked at Intercom is that she would say, “Words mean things.” And then in the knowledge graph world, Dean Allemang has famously said a number of times in, he wrote the book Semantic Web for the Working Ontologist, and he talks a lot about, he says this all the time, “Say what you mean and mean what you say.” And then all of this stuff is just a way of codifying and capturing that so you can do more with it. Is that an accurate summary?

Katariina:
Yes. I mean, at large it reminds me of what the humanities do. As technologists, we’re frowning about humanities. They’re just reading history books and writing about history and it’s very inaccurate and stuff. But they actually are basically… they’re building, I mean, history is about building our memory and that’s the things we choose to remember. There are many things that fall through the cracks and things we don’t choose to remember. So that’s also, it’s not always really good, but we have to write that somehow because we can’t remember everything. So there are certain things we choose to remember, that’s a humanities practice. And there’s certain things in which the way our ethics is and the things we like, and then lawmaking is also things we choose to allow people to do or actually, I mean, law is not allowing this, just telling the things that you should not do or cannot do.

Katariina:
So all these things are in a big scale, a very important part of humanity, and this conscious human behavior in a group setting, like much larger community setting, which we cannot live without. That’s one part of what makes us human. And I guess I’m very European because I actually believe in legislation and lawmaking. I know it’s a slow practice and it’s a little bit… I mean, law is always lagging behind technology. So there’s more things you can already do with tech than is allowed or the other way around. There’s nastier things you can do with tech and law is not yet there. But still I believe that that slower practice and that practice of history and language exploration, of art studies and all of those things, psychology, everything is really a big part. And so as technologists, we also have a responsibility to take that with us as we’re building artificial intelligence.

Larry:
I love that perspective, and that just ties right back this truism about this podcast that it’s mostly about people. And so returning to the humanities and the social sciences to ground technology work, that’s a win. Hey, Katariina, I can’t believe that we’re already coming up close to time. But I always like to give my guests, before we wrap up, is there anything last that you want to say about knowledge graphs and content or just the LLMs or AI in general before we wrap up?

Katariina:
Yeah, I mean for those who are not like to versed in computer science or don’t know much about artificial intelligence, just be critical about all the hype around artificial intelligence. And also large language models or ChatGPT is not just… it’s just one part of building a larger artificial intelligence. And to me, because there’s so much hype around ChatGPT right now, it’s like, “It’s going to change the world, it’s going to take our jobs, dah, dah, dah.” And all these emotions flying around. I just wanted to tell the non-technology people and also colleagues working in artificial intelligence, that large language morals have pretty much solved one part of the brain, which is the language center. Understanding language, computing with language and then also producing correct language.

Katariina:
But there’s so many other parts that also need to be in place and to really build good technology, we want to combine, we want to have hybrid approaches. We want to combine lots of different technology and ultimately we need to, in a good way, connect the people and empower as many people as possible with our tech. So as technologists, we really have to be careful about if we’re building power structures because one thing about large language models is it’s really power hungry. And I don’t mean power hungry as a power structure, but energy hungry. So it needs a lot of electricity, a lot of power, and that’s not sustainable necessary, it’s also very expensive. So that can mean that we are building these things that are expensive and they’re not reachable for the many people who don’t have that much money. So just be really careful and mindful of the kind of technology that we’re building.

Larry:
I appreciate that context a lot. Thank you. Hey, one very, very last thing, Katariina, what’s the best way for folks to connect with you online if they want to follow your work?

Katariina:
So I’m mostly very inactive online. I write one blog post a year or hold a speech a year and so on. So I think you can find a lot of my talks on YouTube. There’s a few, one nice one where I show like Star Trek to people and then I’ve written a few blog posts. So I sound horrible, but Google me and then you’ll find those YouTube videos. You’ll find my Medium blog posts, you’ll find my two Zalando blog posts from the past. And you can also follow me on Twitter, now known as X, with the handle K-A-T-S-I, Katsi, 111, so triple one. K-A-T-S-I 111. And that’s my handle, that’s my handle on Instagram if you want to follow my foodie stuff and my nature pictures. But if you want to follow my more professional stuff, you can find me there. And I’m also happy to connect with anyone on LinkedIn.

Larry:
Great, well I’ll put all that in the show notes as well. Well, thank you so much, Katariina. I really enjoyed the conversation.

Katariina:
Yeah, thanks Larry for having me, and I can’t believe how quickly 30 minutes flew by.