There is a growing awareness and appreciation for taxonomies as information and knowledge management tools. Taxonomies – structured sets of terms tagged to content – support search, information discovery, browse navigation, news alerts and feeds, content recommendation and personalization, content management workflows, and are a part of a semantic layer. Increasingly, various types of organizations and people in more types of roles are finding the need to createtaxonomies for various uses. But how best to create them?

Generative AI technologies, such as ChatGPT, based on large language models (LLMs), can be used to generate answers, narrative text, summaries, outlines, and code, so it would seem logical to expect generative AI to generate taxonomies, too. Experienced and novice taxonomists alike have been experimenting with ChatGPT, generating taxonomy structures of terms, but there are many limitations.

Challenges of Generative AI for Taxonomy Creation

Creating a full, functional taxonomy, even a “small” single-purpose taxonomy, requires an iterative process involving a number of sub-tasks: analyzing the content, describing search use cases, gathering terms as candidate concepts, organizing concepts in hierarchies, identifying and adding alternative labels to concepts, testing the taxonomy, reviewing and analyzing the taxonomy for improvements, and finally implementation and tagging. Some of these tasks benefit from already traditional AI technologies (not necessarily generative AI and LLMs), such as extracting terms from text and auto-tagging, but generative AI can aid with some of the other tasks. Although generative AI can assist with individual sub-tasks, it cannot execute the entire taxonomy project.

Furthermore, the form of generative AI responses either lack certain taxonomy features or are not easy to read. The text type of responses from generative AI could involve a hierarchical list (with bulleted sub-lists for narrower concept hierarchies), but taxonomies are usually more complex than that. Taxonomies have alternative labels (like synonyms), scope notes, and perhaps definitions for concepts. Some taxonomies have additional “related” non-hierarchical relationships. This additional information cannot easily be included in a hierarchy display. 

In theory, generative AI could generate SKOS RDF code (a standard data model for taxonomies) to include all the features of a taxonomy. The RDF file would need to be imported into taxonomy management software to view and edit easily, but the inevitable data errors complicate importing the file. When I asked ChatGPT for a SKOS RDF taxonomy of 50-150 concepts, I got back a taxonomy of 5 concepts full of data errors. 

More significantly, the more properties and variables (related concepts, alternative labels, definitions) you request to add to the taxonomy, the more likely it is that ChatGPT will generate a taxonomy with internal inconsistencies. I tested ChatGPT to create a “thesaurus” with broader, narrower, related, and use/used from relations, but the generated thesaurus lacked consistent reciprocal relationships at each term, as the terms were derived from different source texts.

For example:

Sales. Narrower Term: Direct Sales

Direct Sales. Broader Term: Sales Channels

This brings us to another deficiency in using ChatGPT to create taxonomies. ChatGPT extracts data from divergent sources on the web. A hierarchical relationship may depend on specific content and context. The context could be inappropriate for the taxonomy you are trying to create, and it could be inconsistent with other sources. 

Many of the sources for ChatGPT “taxonomies” are not really information taxonomies at all (what is intended for tagging and search retrieval), but are just categories or outline headings found in texts. This results in hierarchical relationships that are context-specific and not about the concepts. For example, when asked to create a taxonomy in the subject domain of management consulting, ChatGPT returned these hierarchical relationships, among others: 

• Consulting Skills
  • Analytical Skills
  • Communication Skills

Analytical skills and communication skills are not kinds of consulting skills, so they should not be narrower concepts to it.

Finally, generative AI focuses on getting answers from text/content. This is only half of the picture when it comes to creating taxonomies. As taxonomies serve to connect users to content, they need to be designed to take into consideration both the users and the content. That’s why taxonomy design involves tasks that involve end users: interviews, focus groups, brainstorming workshops, and term list suggestions from subject matter experts.

Generative AI for Sub-tasks of Taxonomy Creation

Using generative AI is more suitable for various sub-tasks of taxonomy creation, rather than creating a full taxonomy all at once.

Suggested narrower concepts

When building a taxonomy from top-down, such as starting with user suggestions, you may want some help brainstorming narrower concepts. Whether you ask ChatGPT to create a “taxonomy” on a specific subject or to just create “narrower concepts” to a given subject, you will get the same result. Some or many will not be correctly narrower concepts. But the suggested incorrect concepts can still get the taxonomist thinking of corresponding correct concepts. 

Organizing concepts into hierarchies

A practical use of generative AI is to structure a flat list of terms into a hierarchy. Large lists of hundreds of terms may be created from automated text extraction tools or from search log reports. These are both good sources for terms/concepts. It can be quite tedious to structure long lists of terms into a hierarchy or even assign them as narrower concepts to an existing structure. Any resulting taxonomy structure from ChatGPT may have some errors, but a skilled taxonomist can quickly identify and fit them.

Suggesting alternative labels

Taxonomy concepts have alternative labels (synonyms) to help match to user search strings or words in texts. Most alternative labels come for search log analysis and term extraction for terms with other names that were chosen to be the preferred label. There could be additional alternative labels, though, that these sources did not pick up but might occur in additional content that will be added in the future. Asking ChatGPT to suggest a list of synonyms for a concept, is a good brainstorming technique, even if the majority of the alternative labels should be rejected as inappropriate. 

Generating SPARQL queries

Given the proper instructions, ChatGPT does an acceptable job of generating code, whether for programming, scripting, or query languages. SPARQL is the query language used for SKOS, the data model most commonly used for taxonomies today and used by most taxonomy management systems. If you want to perform focused analysis of a taxonomy, such as identifying certain combinations of types of concepts with types of relationships, SPARQL is the way to do it. 

Conclusions

• Using ChatGPT/LLM technologies can help with various sub-tasks of creating taxonomies but not for a taxonomy as a whole. The skills of a trained taxonomist are needed to put the pieces together properly. 
• Parts of taxonomies created with ChatGPT still require human expert review to correct and refine. Experienced taxonomists, can identify and rectify ChatGPT’s mistakes.
• Some taxonomy-development tasks, such as obtaining input from users through interviews and focus groups, cannot be done with generative AI, and this is where taxonomy experts  are especially helpful. 
• Using LLMs on your own content and data will provide much better context and consistency of terminology generative AI building of taxonomies. EK can also help with this. Read our recent case study.

Learn more about EK’s taxonomy and ontology consulting services, and how a taxonomy strategy can align with a larger knowledge management or content management strategy, which also require humans to develop. provide. Contact us for more information and how we can be of service to you.