Taxonomies and Ontologies Transforming Knowledge Management
With this paradigm, organizations can “visually edit, in the ontology, the path that an AI agent will follow so that it’s not a probabilistic result,” Clarke maintained. “I don’t want to leave it up to chance. I want to know that it’s going to take the end user down this particular path because it’s defined already.” Examples include assisting chatbots backed by generative models to create and fulfill customer support tickets, a task which can be extrapolated to other facets of customer service. Because bots can be directed to the appropriate documentation to answer a specific request, they can ostensibly read it and act accordingly. “We’ve modeled the support process—this is a really specific case—as an ontology,” Clarke said. “The AI agent then picks that up. It does all the nice conversational aspect, really nice use of language, and people can ask questions in three different languages. But it goes back to the ontology.”
Additional GenAI Involvement
The use case Clarke articulated transforms KM from a passive system of organizing and managing content to actively creating timely—and accurate—action to achieve mission-critical objectives. It also happens to coincide with the current vanguard of enterprise AI and current developments in agentic AI. Nevertheless, there are other ways taxonomies and ontologies can support machine learning models to revamp the underlying utility of KM. Some of these entail curtailing, if not outright eliminating, the propensity of models to hallucinate or embellish, creating inaccuracies to questions. “One of the things that taxonomies do is prevent those hallucinations, and they prevent the bias, or can, but only if you built the taxonomies not from the LLM,” Hlava said. For these applications, taxonomies provide grounded enterprise knowledge for language models to parse to understand business concepts in questions or summarization efforts.
Factor’s Wessel stressed that judgment and context are critical when considering GenAI in the taxonomy realm. Context includes provenance and cultural considerations. A third element to keep in mind is that “LLMs are blind to many things. It’s not magic; it’s math. AI does not understand meaning. It relies on mathematical predictions. This erases the cultural context and nuance that may be hugely relevant.” He added that GenAI is not really generative. It’s actually derivative, based on whatever is in its training set. Thus, LLMs are not the best choice for emerging domains.
Lack of standardization in meta-data describing scientific research experiments is a chief KM issue for Mark Musen, Stanford medicine professor of biomedical informatics research. “What you want in the future is to have agents that are on the web that will tell you what experiments you should be looking at, summarize results of experiments, and tell you how to use existing experiments to create new protocols so you can do new experiments,” Musen explained.
Solving the Data Annotation Challenge
There are several obstacles to realizing this future, which would be nothing short of transformative for mastering the data annotation challenge in biomedical science. The problem is that standardized taxonomies and ontologies are not routinely used to describe research experiments or their results. Consequently, search capabilities are compromised, and there’s a dearth of automated measures for linking experiments to existing knowledge. Taxonomies and ontologies can revolutionize this branch of the biomedical industry in two key ways.
Companies and Suppliers Mentioned