Conceptual Spaces

• Conceptual Spaces are geometric frameworks that model concepts as convex regions in a multidimensional space defined by psychologically meaningful attributes like hue, shape, and weight.
• They employ weighted Euclidean and Manhattan metrics to quantify similarity, typicality, and prototype effects, supporting both cognitive science and machine learning applications.
• Extensions using star-shaped and fuzzy set approaches enhance representation by capturing domain correlations and enabling compositional semantic operations.

A conceptual space is a geometric framework for representing knowledge and cognition, in which concepts are modeled as convex regions within a space whose axes correspond to psychologically meaningful quality dimensions such as hue, weight, or shape. This approach, introduced by Peter Gärdenfors, serves as an intermediate representational level—lying between subsymbolic (e.g., neural networks) and symbolic (e.g., logic-based) cognitive architectures—and is characterized by its ability to encode similarity, typicality, prototype effects, and compositional structure with direct geometric and metric semantics (Lieto et al., 2017).

1. Mathematical Structure of Conceptual Spaces

A conceptual space consists of a finite set of “quality dimensions” (e.g., hue, roundness, sweetness), each corresponding to a continuous perceptual or functional attribute (Lieto et al., 2017, Tull et al., 2023). These dimensions are typically grouped into domains (e.g., the color domain comprising hue, saturation, and brightness).

For $\mathcal{D}$ the set of dimensions, an instance $x$ is a point in $\mathbb{R}^n$ (with $n = |\mathcal{D}|$). Domains $\Delta = \{\delta_1, \ldots, \delta_k\}$ partition $\mathcal{D}$ such that each $\delta \subseteq \mathcal{D}$ corresponds to a set of tightly related features.

Distances within a domain are given by a (possibly weighted) Euclidean metric:

$$d_E^\delta(x, y; W_\delta) = \sqrt{\sum_{d\in\delta} w_d (x_d - y_d)^2}, \quad \sum_{d \in \delta} w_d = 1, \; w_d > 0$$

Overall similarity is computed as a weighted sum of intra-domain distances (Manhattan over domains):

$$d_C^\Delta(x, y; W) = \sum_{\delta \in \Delta} w_\delta d_E^\delta(x, y; W_\delta), \quad \sum_{\delta \in \Delta} w_\delta = |\Delta|$$

Concepts are depicted as (often fuzzy) convex or star-shaped regions in this space (Bechberger et al., 2017, Bechberger et al., 2017).

Key properties include:

• Convexity: Each concept corresponds to a convex subset $C \subseteq \mathbb{R}^n$; for all $x, y \in C$ and $0 \leq \lambda \leq 1$, the convex combination $\lambda x + (1-\lambda)y \in C$ (Lieto et al., 2017).
• Prototypes: The central points (usually centroids) of these regions represent the prototype of the category; typicality of a member is a function of its distance to the prototype (Lieto et al., 2017, Bechberger et al., 2017).
• Similarity: An exponential decay function of metric distance ($\exp(-c d(x, y))$) captures the graded similarity between instances (Lieto et al., 2017, Bechberger et al., 2017).

2. Formalizations and Extensions: Convexity and Correlations

While traditional Gärdenfors-style conceptual spaces are defined by convex regions (axis-aligned hyperrectangles in the Manhattan-Euclidean metric setup), several extensions have addressed limitations:

• Star-shaped sets: To encode domain correlations (e.g., age-height dependencies), convexity can be relaxed to star-shapedness. Star-shaped regions are unions of overlapping cuboids (axis-parallel boxes) sharing a nonempty core, enabling representation of “diagonal” or correlated shapes (Bechberger et al., 2018, Bechberger et al., 2017).
• Fuzzy sets: Membership functions allow graded category boundaries. In the implementation of fuzzy star-shaped sets, for a region $S$ and parameters $\mu_0$, $c$, and $W$, the membership of a point $x$ is

$$\mu_S(x) = \mu_0 \max_{y \in S} \exp\left( - c \cdot d_C^{\Delta}(x, y; W) \right)$$

(Bechberger et al., 2017).

• Operations: Efficient algorithms are defined for set-level operations (intersection, union, projection), including explicit formulas for subsethood, concept size, implication, similarity, and betweenness (Bechberger et al., 2017, Bechberger et al., 2018).
• Combinatorial representations: Concept lattices via Formal Concept Analysis provide alternative, discrete structures for organizing concepts, as in the modeling of prosthetic arm functionalities (Ishwarya et al., 2018).

3. Learning and Extracting Conceptual Spaces

Empirical derivation of conceptual spaces can proceed via:

• Psychological similarity data + MDS: Human similarity judgments are collected and multidimensional scaling is used to embed instances in a low-dimensional Euclidean space, producing psychologically valid axes (Bechberger et al., 2018).
• Latent variable models: Recent work grounds the quality dimensions in latent spaces learned by neural networks, such as InfoGANs or VAEs. Under suitable constraints or architectural choices, axes of latent space correspond to interpretable conceptual dimensions, and concepts are convex clusters (Bechberger et al., 2017, Bechberger et al., 2018).
• LLMs and embeddings: Entity embeddings (e.g., from Wikipedia/Wikidata) and LLM-derived spaces with prototype-based axes yield automatically induced conceptual spaces. Features are directions in embedding space; concepts emerge as convex regions, and prototype induction or scalar projections recover concept axes and typicality (Jameel et al., 2016, Chatterjee et al., 2023, Kumar et al., 23 Sep 2025).

Approach	Dimensionality Source	Concept Region Type	Reference
MDS on psychometrics	Human judgments	Convex, fuzzy	(Bechberger et al., 2018)
Latent codes (InfoGAN, VAE)	Neural generative models	Convex clusters	(Bechberger et al., 2017)
Prototype-based LLMs	LLM text embeddings	Prototype-aligned	(Kumar et al., 23 Sep 2025)
Entity embeddings	Distributional statistics	Convex subspaces	(Jameel et al., 2016)

4. Compositionality and Categorical Structure

The compositionality of conceptual spaces has been made precise by categorical semantics, notably via:

• Convex Relations (ConvexRel): The category where objects are convex algebras and morphisms are convex relations. This categorical structure enables grammatical composition of concepts (e.g., combining adjectives and nouns), preserving convexity at each stage (Bolt et al., 2016, Bolt et al., 2017).
• Functorial mapping: Syntax types (e.g., pregroup grammars in NLP) are mapped to semantic spaces (ConvexRel), where compositional rules correspond to morphism composition, ensuring intuitive combinations (e.g., intersective adjective applies as intersection of convex regions).
• Process-theoretic models: Extension into monoidal categories enables both classical (Conv) and quantum (CPM(FHilb)) settings. Concepts correspond to effects, instances to states; conceptual spaces admit both classical and quantum correlation modeling (Tull et al., 2023).

5. Applications, Empirical Evaluations, and Reasoning

Conceptual spaces support a spectrum of cognitive and AI functionalities:

• Explainable AI: Cognitively meaningful axes enable interpretable models for predicting properties such as taste, size, or emotion. LLM-derived spaces with prototype axes support both scalar and ranking-based predictions aligned with human judgments (Kumar et al., 23 Sep 2025, Chatterjee et al., 2023).
• Concept learning and induction: Convex regions support prototype-based induction: positive examples define convex areas; new instances are ranked or classified by proximity. Empirically, automatically learned conceptual spaces have yielded competitive performance in numeric ranking, analogical inference, and link prediction (Jameel et al., 2016).
• Computation of relations: Explicit measures of subsethood, similarity, betweenness, and implication have been formalized and implemented, enabling fine-grained ontologies, analogy, and conceptual mapping (Bechberger et al., 2017, Bechberger et al., 2018).
• Sequential/temporal abstraction: Extensions to modeling temporally unfolding, goal-directed abstract concepts (e.g., chess strategies) treat strategies as convex regions in multidimensional feature space, with sequential trajectories enabling recognition and evolutionary learning of concepts (Banaee et al., 29 Jan 2026).
• Symbolic/subsymbolic integration: By identifying conceptual spaces as a lingua franca, the approach bridges symbolic reasoning, subsymbolic neural representations, and diagrammatic models, supporting hybrid cognitive architectures (Lieto et al., 2017).

6. Theoretical Status, Limitations, and Open Directions

Conceptual spaces serve a primarily epistemic and organizational role: they systematize relational, perception-based, and cognitive information in a metric, multidimensional geometric framework rather than as ontological substrates (Vassallo, 13 May 2025). They:

• Organize similarity and classification without reifying the space itself as a physical or metaphysical arena.
• Support modal reasoning: Only one trajectory through conceptual space may be actualized, aligning with the epistemic rather than metaphysical reading (Vassallo, 13 May 2025).
• Enable theory-driven extensions: Categorical, quantum, and compositional generalizations permit new forms of concept modeling.
• Face ongoing challenges: Automatic discovery of the “right” quality dimensions, interpretability of learned axes, scaling to high-dimensional or dynamic domains, and correlational complexity remain active areas of research (Bechberger et al., 2017, Chatterjee et al., 2023, Kumar et al., 23 Sep 2025).

In sum, conceptual spaces provide a powerful, rigorously formalized, and empirically validated geometric approach to modeling the structure of concepts, enabling compositionality, typicality, prototype effects, analogical inference, and integration across representational levels in both cognitive science and artificial intelligence (Lieto et al., 2017, Bechberger et al., 2018, Bechberger et al., 2017, Bechberger et al., 2017, Bolt et al., 2016, Vassallo, 13 May 2025).