🧬 biology

Characterizing Human Semantic Navigation in Concept Production as Trajectories in Embedding Space

This paper introduces a framework that models human concept production as navigation trajectories within embedding space, utilizing geometric and dynamical metrics to distinguish between clinical groups and concept types across multiple languages while offering a computationally efficient alternative to traditional linguistic analysis.

Original authors: Felipe D. Toro-Hernández, Jesuino Vieira Filho, Rodrigo M. Cabral-Carvalho

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Felipe D. Toro-Hernández, Jesuino Vieira Filho, Rodrigo M. Cabral-Carvalho

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: Mapping the Journey of Your Thoughts

Imagine your brain is a massive, invisible library. When you are asked to name things (like "animals" or "things you can eat"), you don't just pull a random book off the shelf. You navigate through this library. You might start at the "Mammals" aisle, grab a "Dog," then walk over to "Cats," then maybe jump to "Lions."

This paper asks a simple question: What does the path of your thoughts look like?

The researchers built a tool to turn your list of words into a map. Instead of just counting how many words you said, they tracked the movement between each word to see if your thinking process was smooth, chaotic, fast, or stuck in a loop.

The Magic Tool: The "Thought Compass"

To do this, the researchers used AI models (called "Transformers") that understand how words relate to each other. Think of these models as a giant 3D map where words are points in space.

"Dog" and "Cat" are close together.
"Dog" and "Airplane" are far apart.

When you say a list of words, the AI draws a line connecting them. This line is your Semantic Trajectory.

They then measured this line using five "physics-like" tools, similar to how a car's dashboard measures speed and direction:

Distance to Next (The "Jump"): How far did you have to travel in the library to get from one word to the next?
- Small jump: "Dog" → "Cat" (You stayed in the same aisle).
- Huge jump: "Dog" → "Toaster" (You ran across the whole library).
Velocity (The "Speed"): How fast were you moving between ideas?
Acceleration (The "Turn"): Did you change direction suddenly?
- Low acceleration: You were cruising smoothly down the "Pets" aisle.
- High acceleration: You were zig-zagging wildly, switching from "Pets" to "Cars" to "Fruit" in a split second.
Entropy (The "Chaos Meter"): How predictable was your path?
- Low entropy: You were very organized (Dog, Cat, Bird, Fish).
- High entropy: Your path was random and hard to guess (Dog, Toaster, Moon, Pizza).
Distance to Centroid (The "Center of Gravity"): How far did you wander from the "center" of the topic?
- Close to center: You stayed focused on the main idea.
- Far from center: You wandered off into weird, unrelated territory.

What They Found: The Stories the Maps Told

The researchers tested this on four different groups of people and tasks. Here is what the "maps" revealed:

1. The Neurodegenerative Group (Parkinson's & Dementia)

The Analogy: Imagine a driver who is trying to drive to the grocery store but keeps swerving, hitting the brakes suddenly, and taking erratic turns, even though they are staying in the same neighborhood.
The Result: Patients with Parkinson's and dementia showed high acceleration and high entropy. Their paths were "jittery" and unpredictable. They couldn't maintain a smooth flow of thought. However, they also stayed closer to the center (low distance to centroid).
Why? It's like they were stuck in a small, tight circle, unable to explore the wider library, but their movement within that circle was chaotic. This suggests their "executive control" (the part of the brain that plans the route) is struggling.

2. The Swear Word Challenge

The Analogy: Imagine a driver who is allowed to say any word, but they choose to say only curse words.
The Result: When people listed swear words, their paths were fast, chaotic, and full of huge jumps.
Why? Swear words are a unique "neighborhood" in the brain. They are all clustered together (close to the center), but the brain accesses them in a wild, unpredictable way. It's like a chaotic dance party in a small room.

3. The Multilingual Test (Italian & German)

The Analogy: Imagine two people navigating the same library, but one speaks English and the other speaks Spanish.
The Result: Even though the languages were different, the shape of the paths was surprisingly similar across different AI models.
Why? This suggests that the way humans organize thoughts is universal, regardless of the specific language or the specific AI tool used to measure it.

Why This Matters

1. It's a New Kind of "X-Ray" for the Brain
Traditionally, doctors look at how many words a patient says. This paper says, "Wait, how they say them is more important." By looking at the shape of the thought path, we can detect brain issues (like Parkinson's) earlier and more accurately than before.

2. It's Automatic and Fast
Usually, analyzing these lists of words requires a human to sit down and manually categorize every single word (e.g., "Is 'dog' a pet? Is 'cat' a pet?"). This is slow and expensive. This new method uses AI to do the math automatically, making it cheap and scalable.

3. It Bridges Humans and AI
The study showed that different AI models (OpenAI, Google, Qwen) all agreed on the results. This means the "geometry" of human thought is real and consistent, not just an artifact of one specific computer program.

The Bottom Line

This paper treats human thinking not as a static list of words, but as a dynamic journey. By measuring the speed, turns, and chaos of that journey, we can understand how our brains work, spot when they are getting sick, and even see how different languages and AI models perceive the world.

It turns the abstract concept of "thinking" into a visible, measurable path, proving that how we move through our ideas is just as important as the ideas themselves.

1. Problem Statement

Semantic retrieval is traditionally modeled in cognitive science as a "foraging" process involving clustering (exploitation of a semantic subcategory) and switching (exploration of a new subcategory). However, existing methods often rely on labor-intensive, heterogeneous manual pipelines that hinder cross-study comparability. Furthermore, these binary models (cluster/switch) fail to capture the fine-grained, step-by-step dynamics of how humans navigate semantic space.

The authors propose a paradigm shift: viewing semantic retrieval not as a series of discrete categories, but as a continuous trajectory through a high-dimensional embedding space. The challenge is to quantify this navigation using computational metrics that capture both scalar (distance) and directional (velocity/acceleration) aspects of the search process, while accounting for the cumulative nature of human memory (where current thoughts depend on the entire history of the search).

2. Methodology

Data Sources

The framework was evaluated on four diverse datasets spanning multiple languages and tasks:

Neurodegenerative (Spanish): 76 participants (Parkinson's, bvFTD, Healthy Controls) performing a property listing task (PLT).
Swear Fluency (English): 274 participants generating items across categories (Animals, Letters, Swear Words).
Italian & German Property Listing: 69 Italian and 73 German students generating properties for 50 concrete concepts across 10 categories.

Core Framework: Cumulative Semantic Trajectories

Unlike standard approaches that embed words independently, this framework constructs cumulative embeddings.

Trajectory Construction: For a sequence of $N$ items, the embedding $x_t$ at step $t$ represents the cumulative context of items $1$ through $t$ (e.g., $x_2$ encodes "cat dog").
Rationale: This captures the dependency of current semantic states on the full search history, modeling the working memory and inhibitory control required for semantic navigation.

Geometric and Dynamical Metrics

The authors extract five physics-inspired metrics to characterize the trajectory:

Distance to Next: Cosine distance between successive cumulative embeddings. Measures the magnitude of the "semantic jump."
Velocity: The vector difference between consecutive embeddings ( $v_t = x_{t+1} - x_t$ ). Captures direction and speed of movement in semantic space.
Acceleration: The change in velocity vectors ( $a_t = v_{t+1} - v_t$ ). Quantifies the stability of the search; high acceleration implies erratic switching or unstable exploitation.
Entropy: Approximate Shannon entropy of the binarized "Distance to Next" sequence. Measures the predictability and variability of the search pattern.
Distance to Centroid: The cosine distance of each point to the centroid of all unique property embeddings for that participant/concept. Measures global dispersion relative to the semantic center.

Experimental Setup

Embedding Models: Primary results use OpenAI's text-embedding-3-large. Comparisons include Google's text-embedding-004, Qwen3-Embedding-0.6B, and a non-cumulative fastText baseline.
Statistical Analysis: Generalized Linear Mixed Models (GLMMs) with participants and concepts as random factors.
Robustness Checks: ZCA-whitening was applied to test sensitivity to embedding anisotropy.

3. Key Results

Clinical Differentiation (Neurodegenerative Dataset)

The framework successfully distinguished Healthy Controls (HC) from clinical groups (Parkinson's and bvFTD):

HCs exhibited lower Distance to Next, Velocity, and Acceleration, indicating stable, focused semantic clusters.
Patient Groups showed higher values for these metrics, indicating volatile, erratic "semantic jumps" and a failure to maintain stable clusters (executive dysfunction).
Counter-intuitive Finding: Patients showed lower Distance to Centroid than HCs. This suggests that while their search was erratic (high velocity/acceleration), it was spatially constricted to a narrower semantic neighborhood, likely due to a loss of rich sensorimotor traces.

Concept Type Differentiation (Swear Fluency)

Swear Words elicited the highest Distance to Next, Velocity, Acceleration, and Entropy, indicating a highly variable and unpredictable search strategy within a constrained lexical space.
Animals showed the lowest values, reflecting structured, taxonomic exploitation.
Distance to Centroid was lowest for swear words, confirming they occupy a compact, central semantic region despite high internal variability.

Cross-Linguistic and Model Robustness

Language Effects: The Italian and German datasets revealed that kinematic patterns (velocity/acceleration) consistently exposed structural differences between semantic categories (e.g., "Bird" vs. "Vehicle"), though specific category effects varied by language, reflecting cultural and linguistic encoding differences.
Model Independence: Metrics derived from OpenAI, Google, and Qwen models were highly correlated (Pearson $r > 0.75$ for kinematic metrics), suggesting the observed dynamics are inherent to human cognition rather than artifacts of specific model architectures.
Cumulative vs. Non-Cumulative: Cumulative embeddings outperformed non-cumulative baselines for longer trajectories (Neurodegenerative, Swear Fluency), capturing broader semantic patterns. Non-cumulative approaches performed better for very short sequences (Italian/German PLT) where context was limited.

4. Key Contributions

Trajectory-Based Framework: Introduces a novel method to model semantic retrieval as a continuous dynamical system in embedding space, moving beyond binary clustering/switching metrics.
Cumulative Embedding Strategy: Demonstrates that conditioning embeddings on the full search history is crucial for capturing the dynamics of human memory and executive control.
Physics-Inspired Metrics: Defines a set of geometric and kinematic measures (velocity, acceleration, entropy) that provide granular insights into the process of retrieval, not just the output.
Scalability & Automation: Offers a fully automated pipeline requiring minimal human intervention, capable of distinguishing clinical groups and semantic categories across multiple languages and embedding models.
Cross-Model Validation: Proves that semantic navigation dynamics are robust across different transformer architectures (causal vs. bidirectional) and training pipelines.

5. Significance and Future Work

Clinical Applications: The framework provides a quantitative tool for early detection and stratification of neurodegenerative diseases (e.g., Parkinson's, FTD) by identifying specific kinematic signatures of executive dysfunction.
Cognitive Modeling: Bridges the gap between cognitive theories of semantic search and modern NLP representations, offering a way to test hypotheses about human vs. artificial cognition.
Artificial Cognition: Establishes a pipeline for assessing how Large Language Models (LLMs) navigate semantic space compared to humans, potentially revealing gaps in "grounded" cognition.
Limitations: The current assumption of Euclidean dynamics simplifies the anisotropic nature of embedding spaces (though ZCA-whitening showed minimal impact). Future work aims to incorporate non-Euclidean metrics and richer temporal data (speech timestamps).

In summary, this paper establishes a rigorous, mathematically grounded pipeline for quantifying human semantic navigation, demonstrating that the dynamics of the search process (how we move through meaning) are as informative as the content itself, with significant implications for clinical diagnostics and AI alignment.