Humans cannot understand the world by remembering every experience as an isolated event. We group similar things, distinguish differences, assign names, and revise categories when necessary. Concept formation and categorization make it possible to use existing knowledge when encountering something unfamiliar.
Yet categories are not always boundaries written into the world itself. The same object can belong to different categories depending on which characteristics matter and what we are trying to accomplish. Understanding categorization therefore involves questioning the assumptions behind the way we divide things up.
Concepts, Categories, and Classification
These terms overlap in everyday language, but distinguishing them can help:
- Concept: A unit of meaning or knowledge used in understanding and reasoning. Its precise nature remains contested in philosophy and cognitive science.
- Category: A set or class of things treated as belonging together under some perspective.
- Classification: The operation of assigning cases to categories.
- Concept formation: The acquisition or revision of concepts through experience, comparison, and related cognitive processes.
Consider a device described as an “industrial robot,” a “precision positioning system,” or a “substitute for skilled manual work.” The device does not change, but the important properties and possible comparisons do.
Categories Do Not Always Have Clear-Cut Definitions
Some concepts can be defined by explicit rules. Many everyday concepts, however, resist a single defining feature shared by every member.
Ludwig Wittgenstein used the example of “games” to discuss categories connected by overlapping similarities rather than a single common essence. This is the idea of family resemblance.
Eleanor Rosch and colleagues investigated how members of categories differ in typicality. Prototype theory helps explain why people consider some category members more representative than others, even without applying a strict definition.
Nevertheless, categorization cannot always be reduced to similarity. When identifying biological kinds or the functions of artifacts, causal knowledge and background theories may matter more than appearance. Different accounts of concepts explain different aspects of our thinking.
Splitting, Combining, Abstracting, and Reclassifying
| Operation | What it does | Example |
|---|---|---|
| Splitting | Reveals important differences within one category | Divide customers by jobs to be done |
| Combining | Identifies common patterns across differently named cases | Examine shared unmet expectations behind returns, repairs, and inquiries |
| Abstraction | Extracts a relationship or function from a particular object | Move from “pump” to “continuous fluid supply” |
| Reclassification | Reorganizes cases along a different dimension | Switch from product-based to task-based categories |
| Conceptual combination | Combines existing concepts to suggest new possibilities | Combine a distribution network with repair systems |
| Contrast | Uses exceptions and boundary cases to clarify a concept | Ask why one case does not fit the current category |
Abstraction is not merely a loss of detail. By setting aside particular properties, we may identify functions or relationships that can be transferred to other domains. This can be especially useful in cross-domain technology exploration.
Categories Are Not Neutral: An Example
Imagine a manufacturer classifying field failures as “design defects,” “production defects,” or “customer misuse.” This may be useful for assigning departmental responsibility. But it may poorly represent failures arising from unclear instructions, confusing interfaces, and changes in the operating environment.
Reclassifying the same cases by context of occurrence, user objective, earliest detectable symptom, or recoverability may reveal patterns hidden by the original labels.
Categories are useful starting points for analysis, but they are not necessarily the one true description of reality.
Multivariate Analysis and Machine Learning
Clustering, self-organizing maps (SOM), and growing neural gas (GNG) can examine similarity and topological patterns using many attributes or text embeddings. Rather than assigning cases exclusively to predefined labels, these methods can help explore data-driven structures.
Yet a machine-discovered cluster is not automatically a natural category or a real causal entity. The result depends on selected features, distance measures, model parameters, and the composition of the dataset. A two-dimensional projection cannot necessarily preserve all relationships in the original high-dimensional space.
We therefore need to return to original cases and texts, inspect the features that characterize each group, and compare alternative classifications. Finding a structure computationally and assigning it a meaningful interpretation are distinct tasks.
Reconsidering Categories with LLM Assistance
LLMs can articulate properties, suggest category labels, and compare alternative classifications. But quickly assigning a familiar name to something unfamiliar can also suppress the discovery of novel phenomena.
A productive approach is not to request the one “correct” classification. Instead, build several classifications for different purposes and examine the cases that resist each of them. Rare, ambiguous, and exceptional cases may provide clues for forming new concepts.
It is also important to distinguish manual concept mapping, which explicitly represents meaningful relations between concepts, from ConceptMiner’s data-driven concept structure mapping, which explores latent structure in datasets. They are complementary but not identical methods.
Questions Worth Asking
Why was this classification created? Which properties determine category membership? Should a different objective lead to different categories? Which cases do not fit? These questions turn classification from a mechanical task into a possible source of new problems and hypotheses.
Concept formation is not only about placing facts into existing boxes. Sometimes it requires rebuilding the boxes themselves.
References and Further Reading
- Ludwig Wittgenstein, Philosophical Investigations (1953)
- Eleanor Rosch and Carolyn B. Mervis, “Family Resemblances: Studies in the Internal Structure of Categories” (1975)
- Dedre Gentner, “Structure-Mapping: A Theoretical Framework for Analogy” (1983)
- Stanford Encyclopedia of Philosophy: Concepts
- Stanford Encyclopedia of Philosophy: Analogy and Analogical Reasoning