Product Taxonomies are Your Intellectual Property; Why they Deserve More than AI Slop

“Good design is invisible,” a principle that applies as much to digital taxonomy design as it does to the design of physical objects and spaces. At their best, taxonomies guide intuitive user experiences that feel unremarkable because they reflect the user’s understanding and expectations. At their worst, taxonomies can become the source of frustration, creating a barrier between a potential customer and the product they are looking for. Good taxonomies are a competitive differentiator, one that generic solutions simply cannot replicate. They are:

• The language and logic by which your company’s products are presented to the world
  
• A core piece of intellectual property.
  
• An extension of your branding strategy.
  
• The glue that connects customer needs to specific products.
  

Increasingly, platforms such as enterprise knowledge management tools, digital asset management systems, and search/analytic engines offer AI-driven solutions that promise to automate all or parts of taxonomy design. The offer is appealing: good taxonomy design takes time, and AI is promoted as a time and money-saving shortcut. However, the reality is that LLMs and other generative AI tools do not possess the human understanding necessary to produce a quality taxonomy design, especially when those taxonomies will eventually interact with end users in the real world. More importantly, generic AI solutions do not know the nuance of your business or its offerings, and will fall short at representing the tacit knowledge that human experts bring to the table. Taxonomies, particularly those for presenting and navigating your businesses product offerings, are powerful knowledge assets that set the tone for how your business communicates with its customers and the wider world. 

Furthermore, taxonomies support and make organization’s knowledge assets visible, and ensure that the presentation of these assets are aligned to the businesses priorities and IP. Knowledge assets that support product driven businesses might include internal information such as sales volume or inventory levels as well as public facing content such as product images, product information pages, instructions for use, and the metadata that makes up a product description. This is where taxonomies come in, as they can define metadata such as the available values of color variants for a given product SKU, using either generic terms like red, yellow, green, or specific and deliberate branding like brick, sunshine, and meadow. For many of our clients, the language used to describe their unique offerings is well thought out, workshopped, and aligned (sometimes painstakingly) across an organization. Maintaining and reinforcing this understanding is essential to creating and supporting customer experiences that reflect the unique priorities and understanding of a given brand. Taxonomies, a knowledge asset in and of themselves, are also critical semantic structures that support the consistent presentation of additional knowledge assets. In this way taxonomies support alignment internally across a business, and promote a shared understanding between the business and the rest of the world.

In the case of maintaining an organization’s unique verbiage and brand, AI systems have significant limitations when it comes to transitioning from generic language to maintaining a context-specific representation of a domain. Currently, AI systems on their own are not well-equipped to come up with the widely understood community buzzwords that marketing experts know will attract customer interest. Additionally, given the lack of constraints, AI systems may even make more costly mistakes such as utilizing non-preferred language or worse yet, referencing a competitor’s product or brand. Taxonomies on the other hand, are a very simple way to create the direct and explicit correlation of products with preferred descriptions and categorization. Without the guardrails of semantic structure, such as taxonomies, the risk associated with allowing AI systems to interface with customers directly, potentially producing these types of hallucinations, is high. 

Beyond supporting business language and intent, taxonomies provide unique value for companies with a global presence, allowing companies to enforce explicit localization. Language tags on taxonomy terms ensure that terms can be presented in the correct language for a user. Some taxonomies require translation, such as descriptive metadata that should be represented in the language of the intended user (size, color, etc). However, brand names are frequently left untranslated, the original language is the heart of the brand, and would not make sense in a translation. For instance, the computer company Apple, if translated into Spanish as “Manzana” would cause more confusion than clarity. This distinction, between what is logical to translate and what is not, is a unique understanding, and generative AI systems frequently fail to make the correct assumptions about when and how to translate terms. 

So, where do generative AI and taxonomies work well together?

While generative AI alone may not be the best tool for creating new taxonomies all the way through, it can be a useful assistant to taxonomists. Generative AI is very useful for taxonomy design sub-tasks such as writing definitions for labels, suggesting additional broader or narrower concepts, organizing hierarchies, and even generating queries to extract information from the taxonomy. With a “Human-In-The-Loop” to support these workflows, generative AI has proven to be an immensely useful tool for human experts when it comes to taxonomy design, analysis, and development. 

Better yet, taxonomies actually provide an essential input to generative AI systems that can improve the accuracy and contextual relevance of generative AI system responses. By providing structure to content and encoding meaning into knowledge assets, taxonomies support knowledge intelligence, thus supporting explainable AI, and reducing the risks associated with hallucinations. Once primed with a taxonomy of preferred terms, alternative labels, and definitions, generative AI tools can understand the context in which they are operating and are then better equipped to produce content that utilizes the correct language for a given business. In this way, taxonomies serve as a critical link between organizational knowledge assets and customers who wish to access content and products. For text classification or more advanced auto-classification use cases, such as visual analysis, well-specified taxonomies are critical to ensure that products are tagged with actual brand names, preferred labels, or specific color values from an approved list of product colors. 

At EK we have seen the benefit of combining semantic structure such as taxonomies with LLM driven workflows to achieve compelling results. EK enabled a large public health foundation to effectively classify free text survey responses from international participants by utilizing SKOS taxonomies alongside an LLM driven classification approach. Analyzing these survey responses supported injecting customer needs and opinions earlier in the product development pipeline, ensuring that upcoming product development initiatives were well poised to meet target user needs. EK supported the client in creating taxonomies to help understand the localized language of survey participants. These taxonomies allowed the team to classify qualitative free-text responses into clearly defined concept classes, supporting trend analysis across a large data set, while also ensuring that participants were able to use the language that made sense to them. The taxonomy structure allowed us to correlate phrases used in different geographical locations with the same meaning; in this way, we could see that the intention of the response was the same, even if the specific phrasing differed. The hierarchical structure of the taxonomy supported analysis of survey results at different levels of specificity; this solved a previous barrier to efficiently analyzing this dataset, which had become cumbersome and unwieldy when each concept was specified on its own without any categorical groupings. On its own, the LLM would not necessarily have been able to make the connections between different phrases used in different communities, or consistently created the associations between nicknames for a medical treatment and the preferred label for that same treatment. The use of the human-developed taxonomy (and ontology in this case) increased the relevancy of the LLM driven classification and ensured that the community’s understanding and needs were centered in our classification approach. The taxonomy created through this process was both an effective tool that enabled enhanced analysis, and also will also serve as a reusable knowledge asset, able to be utilized to support related research and jumpstart necessary processes for similar work at the organization.

While generative AI may not be able to design your product taxonomy from the ground up, it can be utilized to speed up the tedious parts of the design process, and once available for use, taxonomies are an essential knowledge asset that contribute to improving responses and focusing AI tools on the specifics of your business.

If you want to get started designing taxonomies to support business goals or AI solutions, contact us today.