Why Knowledge Management Architecture Matters, Even When You Think It Doesn’t

Lauren Trees's picture

I talked to Seth Earley, CEO of Earley Information Science, about the role and future of knowledge architecture in machine learning and artificial intelligence applications.

Seth Earley will be a breakout session speaker at APQC’s 2017 Knowledge Management Conference April 27-28. You can learn more about APQC’s 2017 KM Conference here.

APQC: Why do many artificial intelligence experts dismiss the role of knowledge architecture and KM in AI?

Seth: If you think of AI as being only about machine learning and automated processing of large amounts of data, you will quKnowledge Management Role in Artificial Intelligence and Machine Learningickly discover that only works for certain kinds of AI applications: Those where the machine can learn as a result of finding patterns through many iterations across large bodies of data. Self-driving cars come to mind. But if the application needs to talk to humans and give them answers to their specific questions (think IBM’s Watson), that is not a scenario in which you can afford to let the machine learn through mistakes. For that you need to build in context, structure, and curation, which is just another way of saying “knowledge architecture.” Which is precisely what IBM and others do when they discuss “training the AI” to ensure their systems deliver the right answers.

Some AI experts believe that knowledge architecture approaches are outdated, but they are considering problems where knowledge architecture may not be appropriate—such as when attempting to program responses to complex problems that defy traditional programming techniques. Few will argue that a domain model (a knowledge architecture artifact) will not improve the performance of a learning algorithm.

APQC: Can you explain why you believe knowledge architecture is part of every KM solution, even when technology vendors say it’s not?

Seth: The fact is, no matter what the machine learning tool is doing—feature extraction, pattern recognition, even unsupervised—there is still a classification process that needs to take place in order for the system to know what it has found. So while the classification processes or rules may be baked into the “innards” of the machine learning algorithm, they are still in there.

APQC: What is the key to understanding how current KM approaches fit into the AI technology landscape?

Seth: The key thing is to realize that the very frameworks and structures (domain models, metadata models, ontologies, thesauri, etc.) that make KM applications fit for human consumption also make machines learn better. Particularly when you are talking about robots that are expected to talk to humans (like intelligent assistants or chatbots), the same processes for training people apply to training the robots.

APQC: How can KM professionals continue to support current user needs while developing the infrastructure for next-generation KM?

Seth: See the answer to question #3. Also, realize that there isn’t now, nor will there be, some magic “AI” black box that will supersede the need for quality content and data hygiene. When you are working to build a system that is going to help humans make better sense of information, you are simultaneously paving the way for that information to make sense to machines. There are things that need to be done either way: building out ontologies, content models, thesaurus models, and defining your gold standard for content in the organization. And if you are banking on a magic box at some point down the road allowing you to skip the hard work of organizing your stuff now, it is worth noting that your competitors may not be so short-sighted.

APQC: What is the best approach for people without a deep background in AI to educate themselves and get better at interpreting vendor hype in this area?

Seth: This is a tough one because while there is a lot of info out there, much of it is either too high level or too far in the weeds. What is most important is to understand the core principles, like: what is supervised vs. unsupervised learning, reinforcement learning, regression, clustering, and so on. Just having some idea that these concepts exist, even without fully understanding the details, will empower you to recognize when a vendor is blowing smoke. Vendors should be able to explain where their data sources originate, the level of quality expected, the structures involved, how context applies, and what work needs to be done by the business to get value from the solution. If they can’t do that for you, then walk away.


Be the first to comment!