💬 NLP

Introducing multiplex semantic networks as multifaceted representations of creative associative knowledge across multilingual samples

This study demonstrates that multiplex semantic networks, constructed from six cognitive tasks across a diverse multilingual sample, provide a more comprehensive and non-redundant representation of the associative knowledge underlying creativity than single-task approaches, significantly improving the prediction of individual creativity scores through machine learning.

Original authors: Edith Haim, Kurt Haim, Roger E. Beaty, Cynthia S. Q. Siew, Massimo Stella

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Edith Haim, Kurt Haim, Roger E. Beaty, Cynthia S. Q. Siew, Massimo Stella

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 mind as a vast, bustling library. Inside this library, every idea, word, and memory is a book. Creativity is the ability to walk through the aisles and pull out two books that seem completely unrelated—like a "fork" and a "cloud"—and realize they can be connected in a new, useful way.

This paper is like a team of librarians trying to understand how different people organize their mental libraries. They wanted to see if there's a specific "map" of how creative people arrange their books compared to less creative people.

Here is the story of their discovery, broken down into simple parts:

1. The Problem: One Map Isn't Enough

Usually, researchers try to understand your mind by asking you to do just one task, like listing as many words as you can think of related to "school." It's like trying to understand a whole city by only looking at one street. You miss the parks, the bridges, and the hidden alleys.

The researchers asked: What if we looked at six different "streets" of the mind at once?

2. The Experiment: Six Different "Mental Games"

They gathered 518 students from four countries (Austria, the USA, Singapore, and Italy) and asked them to play six different word games. Think of these games as different lenses:

The Flash Game: List words quickly (Verbal Fluency).
The Chain Game: Write sentences where the last word of one sentence starts the next (Sentence Chains).
The Free Association Game: Say the first three words that come to mind when you hear a cue.
The Story Game: Write a short paragraph about creativity in science.

They also asked a Robot (AI) to play these same games. The robot was given a "persona" (e.g., "You are a 20-year-old student from Italy") and told to act like a human.

3. The Big Discovery: The "Multiplex" Library

The researchers built a Multiplex Network. Imagine a sandwich.

Layer 1 is the map from the Flash Game.
Layer 2 is the map from the Chain Game.
Layer 3 is the map from the Story Game.

In a normal study, they would smash all these layers into one flat map. But this team kept them as separate layers stacked on top of each other.

What they found:

Different layers tell different stories. The "Flash Game" map looked different from the "Story Game" map. They weren't just repeating the same information; they were showing different parts of the library.
Creative people have unique maps. When they compared the "High Creative" group to the "Low Creative" group, the layers didn't match up. The creative people's mental libraries had a different shape and structure.
Robots are boring. The AI robot's maps looked almost identical whether it was pretending to be "creative" or "uncreative." The robot's library was too uniform. It couldn't mimic the messy, unique, and distinct structure of a human mind.

4. The Prediction: Guessing Creativity with a Calculator

The team then tried to use a computer program (Machine Learning) to guess how creative a person was just by looking at their library maps. They fed the computer 12 clues, such as:

How big is the library? (Number of words used).
How far apart are the books? (Average distance between ideas).
How clustered are the books? (Do ideas group tightly together?).
How does the "energy" flow? (They simulated a "spreading activation," like a spark jumping from one book to another, to see how fast and far it traveled).

The Result:
The computer got better at guessing creativity when it looked at the combined layers (the sandwich) rather than just one layer.

The strongest clues: The size of the library and how the ideas were connected (structural measures).
The surprise clues: The "spreading activation" (how the spark jumped) added extra power to the prediction.
The emotional clues: Interestingly, people who wrote stories with less "anger" and more "surprise" tended to be rated as more creative.

5. The "Idiosyncratic" Secret

One of the most interesting findings was about "weird" answers.

Standard tests usually throw away answers that only one person gives, calling them "noise" or mistakes.
But this study found that the more "weird" or unique answers a person gave (that no one else thought of), the more creative they were rated.
It's like saying the most creative librarians are the ones who pull out the dusty, forgotten books that no one else remembers.

Summary

This paper argues that to understand creativity, we shouldn't just look at one snapshot of a person's thinking. We need to look at a multi-layered movie of how they connect words, tell stories, and jump between ideas.

Creative minds are like complex, multi-layered libraries with unique paths and surprising connections.
Robots (at least the ones tested here) build libraries that look too similar and predictable, regardless of how they are told to act.
The future: By using these multi-layered maps, we can get a much clearer picture of how human creativity works, capturing the rich, messy, and unique way our brains connect the world.

Technical Summary: Introducing Multiplex Semantic Networks as Multifaceted Representations of Creative Associative Knowledge

Problem Statement
Creativity is a complex cognitive ability relying on the organization and retrieval of knowledge from semantic memory. However, current research often relies on single tasks (e.g., verbal fluency) to measure creativity, capturing only a fraction of the underlying cognitive complexity. It remains unclear whether different cognitive tasks access distinct components of semantic memory or provide redundant information. Furthermore, while network science has modeled semantic memory, single-layer networks fail to capture the multi-dimensional interactions (associative, syntactic, emotional) that characterize creative thought. There is a need for a framework that integrates heterogeneous data types into a unified representation to determine if a multiplex approach offers superior predictive power and structural insight compared to single-task models.

Methodology
The study employed a multiplex network framework to analyze data from 518 individuals across four countries (Austria, USA, Singapore, Italy) and a synthetic comparison group generated by AI.

Data Collection: Participants completed six distinct cognitive tasks aligned by topic ("school" and "science"):
1. Verbal Fluency (VF): Generating nouns within a time limit.
2. Woseco Sentence Chains: Writing consecutive sentences linking concepts.
3. Behavioural Forma Mentis (BFMN): Free association with emotional valence ratings.
4. Textual Forma Mentis (TFMN): Writing short narratives with syntactic dependency analysis.
5. Woseco-Textual Forma Mentis (WSC-TFM): Applying TFMN analysis to the full text of Woseco chains.
6. Alternative Uses Task (AUT): Used as the ground truth for creativity scores, rated by two AI systems (CLAUS and OCSAI).
Network Construction:
- Group-level: Semantic networks were constructed for high- and low-creative groups (split by median AUT score) using correlation-based methods (cosine similarity) and Triangulated Maximally Filtered Graphs (TMFG) for fluency data.
- Individual-level: TFMN and BFMN networks were built per participant. For VF and Woseco tasks, individual-level networks were not constructed directly; instead, spreading activation simulations were run on group-level networks (stratified by fluency) to extract dynamic features for individuals.
Analysis Pipeline:
1. Structural Reducibility: Jensen-Shannon Distance (JSD) was used to quantify structural similarity between task layers. Layers were iteratively merged to identify optimal configurations that minimize complexity without losing information.
2. Machine Learning: A Ridge regression model with nested cross-validation predicted individual AUT scores using 12 non-redundant features derived from network structure, emotional valence, and spreading activation dynamics. SHAP values were used for interpretability.

Key Contributions

Multiplex Framework Validation: The study demonstrates that different cognitive tasks capture distinct, non-redundant aspects of semantic organization. Structural reducibility analyses showed that layers could not be fully collapsed without information loss, supporting the necessity of a multiplex approach over single-task models.
Human vs. AI Structural Differentiation: Human networks exhibited significant structural differences between high- and low-creative groups, which were preserved in the multiplex structure. In contrast, AI-generated networks showed near-identical structures regardless of assigned "creativity" levels, highlighting the homogeneity of AI outputs compared to the variability in human semantic memory.
Predictive Modeling: The study successfully applied machine learning to predict creativity scores using network features. It identified that structural measures (e.g., network size, path length) were the strongest predictors, with spreading activation dynamics providing complementary predictive power.
Idiosyncrasy as a Feature: The analysis revealed that a higher proportion of idiosyncratic responses (words unique to an individual and absent from the group network) correlated positively with creativity, suggesting that standard noise-filtering in network construction may inadvertently discard creativity-relevant signals.

Results

Reducibility: Optimal configurations frequently merged the Textual Forma Mentis (TFM) layer with Woseco-Textual (WSC-TFM) layers, indicating structural similarity between text-based tasks. However, other layers (e.g., VF vs. WSC) remained distinct. Crucially, human high- and low-creative layers generally could not be merged, whereas AI layers could be merged across creativity groups.
Network Properties: Human networks were larger, denser, and exhibited higher modularity and clustering coefficients than AI networks. Human networks showed non-random clustering structures, unlike AI networks which approximated random expectations.
Predictive Performance: The machine learning model achieved modest but significant predictive performance ( $R^2 = 0.116$ ) when using the merged TFM+WSC-TFM layer compared to a single TFM layer ( $R^2 = 0.076$ ).
Feature Importance:
- Structural: Larger network size (more diverse vocabulary) and lower clustering coefficients in TFM networks predicted higher creativity.
- Dynamics: Higher stationary activation levels and maximum activation in spreading simulations predicted higher creativity.
- Emotional: Lower anger and higher surprise in text networks predicted higher creativity.
- Idiosyncrasy: A negative correlation was found between the percentage of valid (group-shared) words and AUT scores, meaning more creative individuals produced more unique, non-shared responses.

Significance and Claims
The paper claims that multiplex semantic networks offer a richer, cross-cultural representation of the associative knowledge underlying creativity than single-task approaches. By demonstrating that different tasks capture complementary information, the study argues for a shift in creativity assessment toward multi-task frameworks. The findings suggest that creative cognition emerges from the interplay of semantic memory structure and dynamic retrieval processes. The authors modestly frame the machine learning results as a "proof-of-concept" with modest predictive power, noting that the use of AI raters for ground truth may have constrained variance. The primary contribution is methodological: establishing that a multiplex approach preserves distinct cognitive signatures of creativity that are lost in aggregated or single-layer models, and that these signatures are robust across diverse cultural and linguistic samples.