TL;DR
This study reveals that the way we move our eyes while freely recalling concepts (like numbers, colors, or animals) can reveal the underlying "shape" or geometry of how those concepts are organized in our minds. By analyzing gaze patterns during verbal fluency tasks, researchers found that eye movements mirror the one-dimensional structure of the mental number line and the two-dimensional structure of the color wheel, suggesting gaze behavior is a powerful, non-invasive window into abstract conceptual organization.
What Problem Does This Paper Address?
Neuroscience has long suggested that our brains use spatial-like representations for abstract concepts, akin to how we navigate physical space. This "cognitive map" theory implies that concepts are organized in structured, often low-dimensional, spaces within our minds. However, directly observing these internal conceptual geometries from human behavior has been challenging. While neuroimaging can show brain activity related to these representations, behavioral readouts are less direct. This paper tackles the question of whether spontaneous eye movements, a readily measurable behavior, can serve as a proxy for these internal conceptual structures, specifically their geometric properties. The authors aim to bridge the gap between neurobiological theories of conceptual representation and observable human behavior by demonstrating that gaze patterns can reflect the dimensionality and relational structure of these abstract mental spaces.
Methods
The core methodology involves analyzing spontaneous eye movements during verbal fluency tasks. Participants were asked to randomly generate words from different conceptual categories: numbers, colors, and animals. Crucially, eye movements were recorded using an eye-tracker with high temporal resolution (1000 Hz) during a 500 ms window before each word was spoken.
For each category, the researchers correlated trial-by-trial changes in eye movement (horizontal, vertical, and combined) with the "distance" between the generated concepts in their respective conceptual spaces.
- Numbers: The mental number line was treated as a 1D structure, with distances defined by numerical difference. Eye movements were analyzed for correlation with left-right (horizontal) and up-down (vertical) gaze shifts.
- Colors: A 2D "color wheel" geometry was assumed. Participants first underwent a training phase where they learned color names and then judged pairwise color similarities. Multidimensional Scaling (MDS) was used to reconstruct individual-specific 2D color spaces. Eye movements were then correlated with Euclidean distances in this reconstructed color space.
- Animals: To explore more complex conceptual spaces, various methods were used to define distances between animal words: cosine distance in high-dimensional word embeddings (FastText), semantic cluster jumps, similarity judgments from a large language model (ChatGPT), and importantly, Euclidean distances along the first principal component (PC1) of word embeddings, which was hypothesized to relate to word frequency.
A key analytical technique was correlating the signed difference in conceptual "value" (e.g., number difference, color distance) with the signed change in gaze position. Additionally, they used Leave-One-Out Cross-Validated Linear Discriminant Analysis (LOOCV LDA) to test if small vs. large numbers (or specific color pairs) could be predicted from gaze positions.
Datasets
The study utilized participant-generated data during experimental tasks. No pre-existing large-scale datasets were used for training models in the traditional machine learning sense. However, for the animal category, the researchers leveraged pre-trained word embeddings from FastText and word frequency data from the Subtlex-it database. For semantic similarity judgments in animals, they employed ChatGPT. The color similarity judgments were collected from participants themselves.
Key Insights and Ideas
- Gaze as a Window into Abstract Geometry: The most striking insight is that eye movements aren't just for visual search; they actively reflect the underlying geometric structure of abstract concepts in memory. This suggests a deep integration of spatial and conceptual processing.
- Dimensionality Matters: The study highlights how eye movements align with the dimensionality of the conceptual space. For the 1D number line, horizontal (left-right) eye movements were dominant, mirroring the linear arrangement. For the 2D color wheel, both horizontal and vertical movements contributed, reflecting the circular structure.
- Left-Right Axis for Order, Not Just Space: For numbers, the left-right axis clearly mapped to numerical order (smaller to larger). For animals, this axis was repurposed to reflect differences in word frequency, suggesting the left-right axis can encode transitions or magnitudes in abstract, non-spatial domains.
- Low-Dimensionality is Key for Complex Concepts: When dealing with complex, multidimensional conceptual spaces like animals, eye movements didn't correlate with high-dimensional embeddings or semantic clusters. Instead, they aligned with low-dimensional projections, specifically the first principal component related to word frequency. This implies our brains might simplify complex conceptual landscapes into more manageable, lower-dimensional representations for behavioral output.
- Pre-conscious Gaze Shifts: The finding that eye movements precede word generation (within a 500ms window) suggests these gaze shifts might be part of the cognitive process of retrieving and preparing to articulate concepts, rather than just a reaction to them.
Key Results
- Numbers: Participants' horizontal eye movements significantly correlated with the left-to-right progression of the mental number line. The magnitude of the numerical change also influenced the magnitude of eye movement. LDA could predict whether a participant was about to say a small or large number based on their gaze position with above-chance accuracy.
- Colors: Eye movements significantly correlated with the Euclidean distances between colors in participants' individually reconstructed 2D color spaces. This correlation was significant for bidimensional eye movements, indicating that gaze reflected the 2D structure of the color wheel.
- Animals: Eye movements did not correlate with high-dimensional semantic representations or cluster jumps. However, horizontal eye movements significantly correlated with differences in word frequency, suggesting participants looked left for smaller frequency differences and right for larger ones. This effect was more pronounced than correlations with other semantic dimensions.
Limitations and Future Work
- Generalizability of Animal Findings: The authors acknowledge that the animal results were exploratory and that other dimensions or models might correlate better with eye movements. They also note that the left-right mapping for animals reflected frequency differences, not absolute frequency, and propose alternative interpretations involving cognitive control.
- Ubiquity Across Participants: The effects were not universally strong across all participants, especially in single-subject analyses. Future work should investigate the conditions and individual differences that influence these gaze biases.
- Causality vs. Epiphenomenon: While evidence suggests gaze biases might be causally involved in mental search, further experiments are needed to definitively characterize this relationship, especially across different conceptual domains and tasks.
- Neural Mechanisms: The paper raises questions about the neural underpinnings, particularly the role of the parietal cortex in representing conceptual knowledge and its link to oculomotor control, suggesting future studies should combine eye-tracking with neuroimaging.
- Task Constraints: The observed effects might be specific to the verbal fluency task. Understanding how task constraints influence the manifestation of these conceptual geometries in gaze behavior is an open question.
Summary updated