Taxonomies and Ontologies Transforming Knowledge Management
“The thing about the RDF semantic schema is that it’s self-describing,” Clarke said. “The hierarchy is described in the predicates. Definitions are described, plus associative relationships, things that link a taxonomy to another taxonomy through named relationships.” Ontologies are ideal for solidifying understanding, and exploring relationships in enterprise knowledge. “If you want to explore new relationships and try to figure out things that are not yet known, then you need the ontology,” Hlava said. For example, if a pharmaceutical company “had a drug that was known to be a good analgesic and is used in ways that are known, what about if we match that to some anecdotal or clinical study data to see how that information might be expanded or used in a different context?”
Automating Taxonomies and Ontologies
Some of the AI capabilities Clarke referenced can drastically automate building taxonomies and ontologies. Users can simply ask language models to generate these constructs. For other implementations, generative models can augment existing ontologies and taxonomies by saying, “Here are lots of synonyms you can add to your taxonomy,” Clarke suggested. “Here are brand new concepts that we’re discovering in content that aren’t in your taxonomy but should be.” These methods may be useful to taxonomists. However, there are numerous ways to automate crafting ontologies and taxonomies that predate doing so with AI models. Entity extractors and term discovery mechanisms can automatically pull out entities and concepts from a corpus of documents.
According to Bram Wessel, founder and principal at Factor, an information architecture and taxonomy consulting firm, creating taxonomies should be about things, not strings. This notion reflects a shift in search technology that moves beyond matching keywords (the strings) to understanding the entities, relationships, and intent (things) behind the search queries. Language is ambiguous. A strike means one thing in baseball and something very different for labor unions. “One solution to this is alternative labels to increase the chance that your assets are tagged with high-quality, consistent, and accurate metadata,” added Bob Kasenchak, a Factor information architect and taxonomist.
An additional wrinkle, as Wessel pointed out, is that it’s difficult to get a large language model (LLM) to get an organization to align on what things mean internally. He commented, “I have yet to see an LLM that can get everyone in, say, a marketing organization to agree what a campaign is.”
There’s also tooling, which Hlava referred to as “an automatic hierarchy generation system,” that can provide groupings of terminology specific to an enterprise’s knowledge. Additionally, OWL is helpful for fleshing out conceptual data models, schema, and relationships among taxonomies. There are also industry-specific ontologies that can jump-start an organization’s journey to devising and curating their own subject area models. From this perspective, the work of taxonomists hasn’t changed much with the advent of generative AI (GenAI).
Improving Language Models
One way taxonomies and ontologies are revolutionizing KM is by supplementing language models. This is particularly acute when dynamic agents—bots—are powered by such models and when taxonomies and ontologies effectively counterbalance the probabilistic nature of the latter. As Clarke commented, a caveat of language models is that they’re probabilistic. “Our clients work in areas like medical or other machine-critical functions where ‘probably the right answer’ isn’t good enough,” Clarke noted. Organizations can use frameworks such as semantic knowledge graphs to clarify the actions and the paths for LLM digital agents to take to consistently achieve the desired outcome of a business process involving KM. In such a deployment, the ends of KM—the reason someone is searching for a document—are prioritized more than the information retrieval it enables.
Companies and Suppliers Mentioned