Working with AI in Taxonomy Builder: Human in the Loop
Use of Taxonomy Builder goes far beyond accepting or rejecting AI suggestions. It spans the entire taxonomy building workflow, from setting context and generation parameters to editing outputs and establishing governance policies that keep humans in control.
Main Takeaways
- HITL in taxonomy building is the whole workflow, not a chat conversation — human involvement spans context-setting, parameter choices, editing suggestions, and governance.
- AI handles the mechanics, humans set the boundaries — generating concepts, synonyms, and definitions is delegated to AI so taxonomists can focus on structure, scope, and usefulness of results.s.
- Context is what AI can't figure out on its own — clear and precise human inputs determine the quality of what's generated after.
- Governance keeps humans in charge long-term — a taxonomy governance plan that includes AI generation guidelines, parameter rules, and acceptance criteria ensures the taxonomy stays accurate and useful as it evolves.
Working with AI in Taxonomy Builder: Human in the Loop
In my previous blog post, “How AI and Taxonomy Builder Support the Building of Taxonomies,” I explained how Generative AI in general and the Graphwise Taxonomy Builder feature in particular assists with taxonomy development. I also explained that Generative AI is best used for various taxonomy-creation subtasks, rather than generating a full taxonomy, and these subtasks offer greater opportunity for human review. This blog post explains in more detail the role of human involvement, human review and more, in Taxonomy Builder.
Human-in-the-Loop with AI
Human-in-the-Loop (HITL) refers to an iterative process where humans (subject matter experts, data scientists, or end-users) interact with AI systems to guide, validate, and refine their outputs. The concept of HITL originated with machine learning (ML) systems, in which humans accept or reject ML results and at the same time give feedback to the system so it could improve. The word “loop” is used to refer to the system workflow. There could be more than one point in the workflow in which to provide human interaction and feedback.
Generative AI platforms involve HITL through their chat interface, which allows the human to refine their prompts/queries and create follow-up instructions. The chat interface is suitable for question-based requests, but not for building taxonomies or parts of taxonomies. Taxonomy Builder, thus does not use a chat interface, but rather offers many other points of interaction for the human user, which actually amount to greater human involvement in the taxonomy building process. The “loop” is indeed the entire taxonomy building process and not merely a chat conversation.
If we compare LLM-based technologies, such as Taxonomy Builder, with other AI technologies, such as those based on machine learning, text analytics, and NLP, including corpus analysis (Graphwise Semantic Analytics) and automated semantic tagging (Graphwise Concept Tagging), we find Taxonomy Builder has more human input options than these other AI-based solutions, although that’s largely due to the multi-step nature of building taxonomies.
Providing context
To build an appropriate taxonomy requires understanding of the context of the content, the users, the use case, and the front-end application requirements. LLMs cannot figure this out beyond some of the content context and the full context is too complicated to capture in the instructions of a prompt. The initial HITL activity involves providing context.
This includes naming the project unambiguously, giving a subject and description of the project and creating the start of the taxonomy with the high-level structure of concept schemes. You can also add detailed descriptions and subjects for each of the concept schemes which Taxonomy Builder will take into consideration.
Any additional taxonomy concepts in the branch of the taxonomy where AI generation will occur also provide context. The context of a partially created taxonomy brings information to LLMs as an example of how to build out the taxonomy further.
If generating taxonomy from the high level of the project or concept scheme level, you can also enter additional context information in the “Enter generation instruction…” field in Taxonomy Builder.
Providing additional instructions and generation settings
The context of an example, however, is not sufficient to give all information needed, especially if it is only the start of a taxonomy. Additionally, HITL involvement includes selecting the starting node for generation, providing additional instructions, choosing the taxonomy size, and setting various parameters.
The most important setting is for the taxonomy size. Taxonomy Builder offers the option to select “Small,” “Wide, or “Deep.”
In addition to choosing one of these size types, you can give more specific instructions. The same field “Enter generation instruction…” where you can add context information is also where you can provide further detailed instructions. You can enter instructions for an exact number or range of levels of hierarchy and of concepts per level.
For example: “Generate a total of 25-40 industry concepts in this concept scheme, grouped by categories provided by top concepts into a two-level hierarchy. Do not add a third level.”
You can also enter instructions for an exact number or range of alternative labels to generate per concept (in combination with selecting the settings option of generating narrower concepts).
Additional settings in the user interface include:
- Generation parameters for specified narrower concepts to exclude, terms to include, and terms to include.
- Additional concept attributes of alternative labels and definitions to generate at the same time of concept generation, rather than later.
- Advanced numeric LLM parameters of “temperature,” which controls the randomness or creativity generated by LLMs during inference, and the maximum number of tokens generated, how much information should the model return. These impact the results in ways other than simply size.
Editing generated suggestions
The most basic HITL interaction for any automated system is to accept or reject automated suggestions. Taxonomy Builder offers multiple opportunities to accept or reject suggested concepts, alternative labels and definitions. It also provides the opportunity to edit the suggestions at various steps in the workflow.
When a new taxonomy hierarchy or branch is generated, you can browse the concepts alphabetically, hierarchically, or search them. You can edit their labels, additionally generate alternative labels and definitions for concepts individually or by multi-select batch selection. Individual concepts may be deselected before accepting all, or all concepts may be deselected and then individual concepts may be selected back.
When using the Taxonomy Builder’s “Extend your Taxonomy” feature, multiple narrower concepts, alternative labels, and definitions are generated for a single concept. The generated narrower concepts are not added unless you select/approve them. Generated alternative labels and definitions, on the other hand, will get added unless you deselect them. This is a logical design for HITL, since adding new narrower concepts needs to be justified by the use case, whereas adding alternative labels and definitions merely supports the taxonomy concepts that are already approved.
AI and humans in combination for taxonomy building
AI generation with Taxonomy Builder can be combined with a manually created taxonomy in different ways. It goes beyond mere HITL interaction of setting generation parameters and editing a generated taxonomy. Taxonomy owners or taxonomists may choose to design the upper levels of a taxonomy without AI assistance, but then prefer to use AI to save time with the tedious task of building out detailed concepts, alternative labels, and definitions. On the other hand, subject matter experts may be able to build out a taxonomy in detail, but don’t know how best to structure the start of a taxonomy and need the example of a starter taxonomy, which AI can generate with instructions.
Organizations have often turned to licensing published taxonomies in desired subject domains, to save time and make up for lack of subject expertise. Even these purchased taxonomies need to be edited and adapted to a specific use case. Starting with a generated taxonomy with Taxonomy Builder is similar to starting with a pre-built industry taxonomy, but it’s even more suitable for a specific use case, based on the generation instructions, context information, and settings the user provides.
Taxonomy Builder’s functionality is designed to build a starter taxonomy, to which you must add more details, either manually or generated incrementally for review at each individual concept with the “Extend your Taxonomy” feature. The size setting options in Taxonomy Builder are “Small,” “Wide,” and “Deep” but not both wide and deep together or “full” taxonomy. When generating a full taxonomy hierarchy from any node, Taxonomy Builder provides suitable suggested concepts, if not a comprehensive list of concepts, so the taxonomist knows to manually add more similar sibling concepts.
In other cases, most of a custom taxonomy may be built manually, but AI-generation from Taxonomy Builder is helpful for generating the added details of alternative labels and definitions, whether some or all of them.
Taxonomy governance: The human in charge
Taxonomy governance encompasses the policies, guidelines, and documentation pertaining to a taxonomy’s design and use, especially with respect to its continued maintenance and future updates. A custom governance plan is developed and implemented entirely by humans. But if the taxonomy is at least partially AI-generated, then AI-generation guidelines and settings need to be included in the governance plan.
Taxonomy governance for AI-generation and the use of Taxonomy Builder for a specific taxonomy include policies or guidelines for the following:
- At what levels of the taxonomy to build out with GenAI
- What to set as parameters in the additional generation settings
- When not to accept generated suggestions or to edit them
- The minimum and maximum number of narrower concepts, alternative labels, and definitions to have per concept
- How many levels deep the hierarchies should be
- What to include in additional generation instructions (prompts)
Developing and following a taxonomy governance plan is a way that humans will always remain in control of taxonomies.
The future of taxonomy building
The most important insight from working with Taxonomy Builder is that AI doesn’t reduce the need for human judgment — it relocates it. AI can handle the tedious mechanics — generating concepts, writing definitions, hunting for synonyms. And taxonomists and subject matter experts can focus on what only humans can do: understanding the context, setting boundaries, and deciding whether the taxonomy actually serves its users.
That shift — from execution to governance — is what makes AI genuinely useful in knowledge management work. The loop isn’t a safety net. It’s where the real expertise lives.
Details
How to Build a Business Taxonomy?
Graphwise Graph Modeling helps companies build a business taxonomy with ease and is regarded as a premium taxonomy and ontology management software by industry publishers such as KMWorld and Gartner.
Learn moreFAQ
Any Questions? Look Here
AI-generated taxonomy suggestions can be highly reliable as a "force multiplier" for taxonomists, but they should not be trusted as a total replacement for human expertise. While tools like the Graphwise Taxonomy Builder improve accuracy by grounding Large Language Models in the context of the start of your taxonomy to avoid the generic pitfalls of public AI, full automation often lacks the specific context and hierarchical consistency required for complex enterprise use cases. Consequently, a "human-in-the-loop" approach is essential; you should treat AI suggestions as candidate concepts, synonyms, or definitions that require validation, refinement, or rejection.
In an AI-assisted workflow, a taxonomist transitions from a creator of terms to a "human-in-the-loop" governor who provides essential oversight and strategic direction. While AI tools like Large Language Models (LLMs) accelerate the process by automating repetitive tasks—such as suggesting concept hierarchies, generating synonyms, and drafting definitions—the taxonomist is responsible for validating and refining these outputs to ensure accuracy and contextual relevance. By curating the high-quality semantic structures that serve as the "backbone" for AI systems, taxonomists provide the necessary semantic context to reduce hallucinations and improve the precision of technologies like GraphRAG and enterprise knowledge graphs. Ultimately, the taxonomist ensures that the automated components remain aligned with business goals, standards-compliance, and human intent.
While building a high-quality taxonomy traditionally requires collaboration between a professional taxonomist and a Subject Matter Expert (SME), it is increasingly possible to develop one without a dedicated expert by leveraging modern AI-driven tools and automation. Technologies such as Graphwise’s Taxonomy Builder allow business users and SME to bridge the "expertise gap" by automatically suggesting concept hierarchies, alternative labels, and definitions based on existing content or domain context. However, even with these accelerators, a "human-in-the-loop" approach remains essential for governance and quality control to ensure the resulting structure is logically sound, unbiased, and tailored to the specific business use case.
Taxonomy governance is the set of documented policies, roles, and procedures used to manage and maintain a taxonomy throughout its lifecycle, ensuring it remains accurate, consistent, and aligned with business objectives. To implement it, organizations must first define clear responsibilities for stakeholders — such as taxonomists, subject matter experts, and IT teams — to determine who can request, approve, and execute changes. Implementation also involves establishing standardized workflows for routine updates (e.g., adding or merging concepts) and major structural revisions, supported by comprehensive documentation that covers the taxonomy’s purpose, editorial guidelines, and tagging policies.
In knowledge management workflows, humans should override AI suggestions when high-stakes strategic judgment, domain-specific nuance, or ethical oversight is required. While AI effectively automates routine tasks like metadata extraction and initial concept mapping, human intervention is essential for validating machine-generated relationships to ensure they are accurate, contextual, and deliver actual business value. This "human-in-the-loop" approach serves as a critical quality control mechanism to mitigate risks such as unintended bias, context rot, or hallucinations, particularly in specialized domains where AI may lack deep understanding of systemic causalities. By acting as knowledge stewards rather than mere engineers, humans provide the necessary guidance to align AI outputs with organizational goals, ensuring that the resulting knowledge graph remains a trustworthy, explainable, and strategic asset.