💬 NLP

Entropy in Semantic Memory Navigation in Blind and Sighted Individuals: The Effect of Visual Experience

This study demonstrates that visual experience fundamentally shapes the organization and navigation of semantic memory, as evidenced by distinct entropy patterns in property retrieval between congenitally blind and sighted individuals, where the latter show higher unpredictability for abstract concepts while the former exhibit it for visually salient concrete concepts.

Original authors: Felipe D. Toro-Hernández, Rodrigo Lagos, Sergio E. Chaigneau

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Felipe D. Toro-Hernández, Rodrigo Lagos, Sergio E. Chaigneau

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain's memory as a giant, bustling library where you go to find the "files" on everything you know—from a fluffy dog to the feeling of jealousy. For a long time, scientists have debated how this library is organized. Some say it's built on a foundation of sensory experiences (what things look, feel, and smell like), while others think it's more like a pure text-based database that doesn't need those senses to work.

To settle this, researchers Felipe D. Toro-Hernández, Rodrigo Lagos, and Sergio E. Chaigneau decided to peek inside the libraries of two very different groups: people who have been blind since birth and people who can see. They didn't just ask, "Do you know what a dog is?" Instead, they watched how these groups navigated their mental libraries to list properties of words.

The Game: A Mental "Property Listing" Race
The researchers used a dataset where participants had one minute to list as many things as they could think of for a specific word. They looked at concrete words (like "penguin" or "potato") and abstract words (like "democracy" or "jealousy").

To measure the "chaos" or "predictability" of their answers, the team used a clever tool called semantic entropy. Think of this like measuring the "jitteriness" of a path.

Low Entropy (Smooth Path): If you say "dog," then "bark," then "fur," then "paws," you are walking a very straight, predictable line. You aren't jumping around; you're staying in the same neighborhood of ideas.
High Entropy (Bumpy Path): If you say "dog," then "shark," then "love," then "traffic," your path is zig-zagging wildly. You are jumping all over the map, making it hard to predict where you'll go next.

The Big Discovery: Two Different Maps
The study found that sighted people and blind people use their libraries in surprisingly different ways, even though they both know the words.

The Sighted Experience: For people with sight, the library has a clear split. When they think about abstract concepts (like "jealousy" or "worry"), their paths are very bumpy and unpredictable (high entropy). It's like they are wandering through a foggy, abstract maze because there's no single picture to anchor them. But when they think about concrete things (like "apple" or "zebra"), their paths are smooth and predictable (low entropy). This is likely because they can picture the object, which acts like a magnet, keeping their thoughts on a straight track.
The Blind Experience: For people who have been blind since birth, the map looks different. The study suggests that the "foggy maze" of abstract concepts and the "smooth track" of concrete things are much more similar to each other. In fact, for some concrete items that are famous for being visual—like a penguin—blind participants showed higher entropy (more bumpy paths) than sighted people did.

What This Means for the "Senses" Debate
This finding is a bit of a twist in the story of how we learn.

It rules out the idea that you must have seen something to understand it. Blind participants navigated their memories successfully, proving that you don't need visual experience to have a working library.
It suggests that while you don't need vision to know a concept, having vision changes the architecture of your memory. For sighted people, vision acts like a "traffic controller," keeping thoughts about physical objects on a straight, predictable road. Without that visual traffic controller, blind individuals rely more on language and context for everything, making their paths for concrete objects look a bit more like their paths for abstract ideas.

How Sure Are We?
The authors are careful to say these results suggest a reorganization of the brain's memory map, rather than proving a permanent, unchangeable rule. They measured this using a specific dataset of 48 Italian speakers (22 blind, 26 sighted) and used advanced computer models to track the "jumps" between words.

Interestingly, the study found that for blind participants, the difference between "concrete" and "abstract" was compressed. While sighted people had a huge gap between how they handled "penguins" (smooth) and "democracy" (bumpy), blind people had a much flatter landscape. In some cases, the "penguin" path was actually bumpier than the "justice" path for blind participants.

The Takeaway
This study doesn't say blind people have "worse" memory or that sighted people have "better" memory. Instead, it suggests that our brains are like adaptable cities. If you build a city without visual landmarks, the streets might look different, and the routes you take to get to the "post office" (a concrete object) might be more winding and less predictable than in a city full of visual signs. The destination is the same, but the journey through the mind is uniquely shaped by the senses we use to explore the world.

Technical Summary: Entropy in Semantic Memory Navigation in Blind and Sighted Individuals

Problem and Motivation
The paper addresses a central debate in cognitive neuroscience regarding the representation of semantic memory: whether knowledge is grounded in sensorimotor systems (embodied accounts) or exists in amodal linguistic networks (disembodied accounts). While congenitally blind individuals demonstrate robust language abilities, suggesting that visual experience is not strictly necessary for semantic representation, the extent to which the absence of visual input alters the organization and dynamic retrieval of semantic knowledge remains unclear. Previous research indicates that blind individuals may rely more on linguistic and contextual regularities, potentially compressing the usual distinctions between concrete and abstract concepts. However, traditional static measures (e.g., feature counts, reaction times) often fail to capture the temporal dynamics of how semantic memory is navigated. This study investigates whether the navigation patterns (retrieval dynamics) of semantic memory differ between congenitally blind and sighted individuals, specifically examining how visual experience influences the predictability of retrieving properties for concrete versus abstract concepts.

Methodology
The study utilized a secondary analysis of the "BLIND" dataset (Lenci et al., 2013), comprising 48 native Italian speakers (22 congenitally blind, 26 sighted) matched for age, education, and sociocultural background. Participants performed a property listing task, orally describing the meaning of 50 noun concepts (spanning concrete categories: Bird, Mammal, Fruit, Vegetable; and abstract categories: Emotion, Ideal) within a one-minute window.

To move beyond static feature analysis, the authors employed Natural Language Processing (NLP) techniques to model semantic memory navigation as a trajectory through a high-dimensional semantic space:

Embedding Construction: A transformer-based multilingual embedding model (Google Gemini Embedding) was used to map the sequence of produced features. Crucially, embeddings were constructed cumulatively ( $x_t = E(w_1, \dots, w_t)$ ), where each step represents the semantic content of the current property combined with all previously generated properties for that concept.
Semantic Entropy Calculation: The study computed "semantic entropy" to quantify the unpredictability of the retrieval trajectory.
- Cosine distances between successive cumulative embeddings were calculated to measure semantic "jumps."
- These distances were binarized based on the within-trajectory median to create a binary time series of "high-jump" vs. "low-jump" steps.
- Shannon entropy was calculated on this binary sequence. Higher entropy indicates more variable, less predictable semantic exploration, while lower entropy suggests constrained, regular trajectories.
Statistical Analysis: Generalized Linear Mixed Models (GLMM) with a Gamma distribution were used to assess group differences (Blind vs. Sighted) across semantic categories and specific concepts, controlling for participant and concept variability.

Key Results
The analysis revealed distinct navigation patterns between the two groups, characterized by differences in entropy profiles:

Group-by-Category Interaction: Sighted individuals exhibited significantly higher entropy for abstract concepts (particularly Emotion) compared to concrete concepts. This suggests that retrieving abstract concepts involves more variable and less predictable semantic exploration. In contrast, blind participants did not show this clear distinction; their entropy levels were more uniform across categories.
Concrete Concept Reversal: Notably, for blind participants, certain visually salient concrete concepts (e.g., penguin) exhibited higher entropy than abstract concepts, a pattern not observed in the sighted group.
Granularity of Differentiation:
- Sighted Group: Showed broad, systematic differentiation at the concept level. Abstract concepts like jealousy and worry consistently demonstrated higher entropy than a wide range of concrete items.
- Blind Group: Showed a "flatter" entropy profile with fewer significant distinctions. Differences were localized, often confined to within-category comparisons (e.g., among animal concepts) rather than a global abstract-concrete divide. While cheerfulness showed higher entropy than some concrete items, other abstract concepts showed lower entropy than specific concrete items (e.g., penguin).

Key Contributions

Dynamic Metric for Semantic Memory: The paper introduces and validates an embedding-based semantic entropy metric to quantify the step-by-step granularity of semantic retrieval, offering a more dynamic alternative to traditional clustering or switching metrics.
Reorganization of Semantic Space: The findings demonstrate that while blind individuals possess coherent semantic knowledge, the internal geometry of their semantic space differs from sighted individuals. The absence of visual experience appears to reduce the contrast between abstract and concrete retrieval dynamics.
Role of Visual Experience: The study provides evidence that visual experience is not a sine qua non for semantic processing (as blind individuals navigate semantic space successfully) but is a critical factor in shaping the organization and predictability of that navigation, particularly for concepts typically grounded in visual perception.

Significance and Claims
The authors claim that their results reconcile two opposing strands of literature: the preservation of semantic competence in blindness and the existence of selective representational differences. The study argues that blind and sighted individuals share a broadly similar semantic space but traverse it differently.

The paper posits that sensorimotor systems are not strictly necessary for the formation of semantic representations but play a functional role in structuring how knowledge is accessed. For sighted individuals, visual grounding constrains the search for concrete concepts, leading to lower entropy, while abstract concepts require more exploratory, context-dependent navigation (higher entropy). For blind individuals, the reliance on linguistic and contextual pathways for all concepts (including concrete ones) results in a compressed semantic landscape where the retrieval dynamics of abstract and concrete concepts are more similar.

The authors conclude that these findings support a nuanced view of embodied cognition where sensory experience influences the dynamics of semantic memory without being essential for its existence, challenging both strong embodied accounts (which might predict failure without vision) and strict amodal accounts (which might predict identical dynamics regardless of sensory history).

Technical Summary: Entropy in Semantic Memory Navigation in Blind and Sighted Individuals

More like this