💬 NLP

Chinese sensorimotor and embodiment norms for 3,000 lexicalized concepts

This paper introduces a comprehensive database of 11-dimensional sensorimotor and embodiment norms for 3,000 Mandarin Chinese concepts, demonstrating their high reliability, predictive power in lexical decision tasks, and partial recoverability from purely linguistic representations.

Original authors: Jing Chen, Gábor Parti, Yin Zhong, Chu-Ren Huang, Marco Marelli

Published 2026-05-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jing Chen, Gábor Parti, Yin Zhong, Chu-Ren Huang, Marco Marelli

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

The Big Idea: How Words "Feel" in the Body

Imagine your brain isn't just a library of definitions, but a high-tech simulation center. When you hear the word "apple," your brain doesn't just think "red fruit." It briefly simulates the crunch, the sweetness, the smell, and the motion of biting it. This is called embodied cognition: the idea that our understanding of words is built on our physical experiences.

The researchers behind this paper wanted to map out exactly how this works for Mandarin Chinese. They asked: If we have a word, how strongly does it trigger our senses (sight, sound, taste) and our muscles (moving hands, legs, head)?

The Project: Building a "Sensory Map" for 3,000 Words

The team created a massive database for 3,000 common Chinese words. Think of this as a giant spreadsheet where every word gets a "report card" with 12 grades:

Six Senses: How much do you see, hear, taste, smell, touch, or feel inside (like hunger or pain) when you think of this word?
Five Body Parts: How much do you use your hands/arms, legs/feet, mouth/throat, head, or torso (body trunk) to interact with this word?

They asked 378 native speakers to rate these words on a scale from 0 (no feeling at all) to 5 (a very strong feeling).

The Results so far:

Vision is King: Just like in English, the most common "feeling" for Chinese words is visual. If you think of a word, you are most likely "seeing" it in your mind.
The "Gut" Feeling: A surprisingly large number of words (especially abstract ones like "longing" or "pain") triggered strong internal sensations (interoception).
Action vs. Sensation: Words that make you feel "embodied" (like a whole-body experience) are more closely linked to moving your body (motor actions) than just seeing or hearing things.

Study 2: The "Speed Test" (Do these feelings help us read faster?)

The researchers wanted to know: Does knowing how "physical" a word is help us recognize it faster?

They took the 3,000 words and ran them through a Lexical Decision Task. This is a computer game where you have to press a button as fast as possible to say if a string of letters is a real word or a fake one.

The Metaphor: Imagine your brain is a car. The "sensorimotor ratings" are like the fuel. The researchers tested different types of fuel formulas to see which one made the car go fastest.

The Winner: They found that the best "fuel" wasn't just the strongest single sense (like "how visual is it?"). The best predictor was a complex recipe that looked at the strongest sense but also gave a little boost to the other senses involved.
The Takeaway: Words that feel "physical" and involve multiple senses or body parts are processed faster by the brain. The brain loves a rich, multi-sensory experience.

Study 3: Can a Robot "Feel" Without a Body?

This is the most philosophical part. Large Language Models (AI) like the ones we chat with today are trained only on text. They have never seen a sunset, tasted an apple, or felt a hug. They are "disembodied."

The researchers asked: Can we predict how a human feels about a word just by looking at how the AI "sees" the word in text?

The Experiment:
They fed the AI's word data (mathematical representations of words based on how often they appear together in text) into a simple math model to guess the human ratings.

The Results:

It Works (Sort of): The AI could guess human feelings with about 62% accuracy on average.
The "Vision" Advantage: The AI was very good at guessing how "visual" a word is. This makes sense because language describes visual things very well (e.g., "The red, bright sun").
The "Taste" Problem: The AI was terrible at guessing taste and smell.
- Why? Because it's hard to describe a specific taste in words without actually tasting it. Also, most words in a text corpus aren't about food, so the AI didn't get enough "data" to learn the pattern.
The Geometry: Even though the AI couldn't guess every single number perfectly, it got the relationships right. If humans thought "apple" and "orange" were similar in taste, the AI also thought they were similar. The "shape" of the sensory world was partially recoverable from text alone.

Summary

We made a map: We now have a detailed guide showing how 3,000 Chinese words connect to our senses and muscles.
Physical words are faster: The more a word feels like a real-world experience, the faster our brains recognize it.
Text is a good, but imperfect, proxy: Even without a body, an AI can learn a lot about how humans experience the world just by reading books and news, especially for things we can see. But for things like taste and smell, text just isn't enough to capture the full human experience.

Where to find the data: The researchers made all their data and code public, so other scientists can use this "sensory map" for their own studies.

Technical Summary: Chinese Sensorimotor and Embodiment Norms for 3,000 Lexicalized Concepts

Problem Statement
The paper addresses a critical gap in cognitive science and embodied artificial intelligence research: the lack of large-scale, high-dimensional normative resources for non-Indo-European languages. While understanding how conceptual knowledge is grounded in bodily experience is central to these fields, existing sensorimotor databases are predominantly English-centric or limited in scope for Mandarin Chinese. Previous Chinese resources either covered few words (e.g., 232–689 items), lacked comprehensive dimensionality (e.g., only perceptual modalities), or were restricted to specific grammatical categories (e.g., nouns only). This scarcity hinders empirical investigation into the relationship between sensorimotor experience and lexical processing in Chinese, as well as the ability to train or evaluate large language models (LLMs) on grounded conceptual knowledge.

Methodology
The study is structured into three distinct phases:

Data Collection (Study 1):
- Participants: 378 native Mandarin speakers were recruited via Prolific and Chinese social media platforms.
- Stimuli: A wordlist of 3,000 frequently used lexicalized concepts was assembled by pooling items from four existing Chinese norming studies and adding 429 high-frequency words from the Sinica Corpus 4.0 and Chinese Gigaword 5.
- Ratings: Participants provided ratings on two scales:
  - 11-Dimensional Sensorimotor Ratings: Following the Lancaster framework, participants rated words on six perceptual modalities (Visual, Auditory, Gustatory, Olfactory, Haptic, Interoception) and five action effectors (Hand/Arm, Leg/Foot, Head, Mouth/Throat, Torso) on a 0–5 scale.
  - Unidimensional Embodiment Ratings: Participants rated the strength of bodily experience evoked by each word on a 0–6 Likert scale.
- Quality Control: Rigorous attention checks and a Chinese language competence mini-test were administered to ensure data validity.
Lexical Decision Validation (Study 2):
- Dataset: 2,044 words from the sensorimotor norms were matched with lexical decision data (response times and error rates) from the MELD-SCH database.
- Predictors: Nine composite variables were derived from the sensorimotor ratings to predict lexical performance:
  - Two variants of Perceptual Strength of Embodiment (PSE): PSE-Perceptual (5 modalities) and PSE-Sensorimotor (all 11 dimensions).
  - Six established metrics from Lancaster norms: Maximum strength, Minkowski-10, Minkowski-3, Euclidean distance, Summed strength, and PCA component.
  - The unidimensional Embodiment rating.
- Analysis: Bayesian linear regression models were fitted to predict standardized response times (zRT) and error rates (ERR), controlling for word length and frequency.
Linguistic Recovery Analysis (Study 3):
- Approach: The study investigated the extent to which sensorimotor ratings can be predicted from purely linguistic representations (static word embeddings from Tencent AI Lab) using ridge regression.
- Metrics: Predictive accuracy was measured via $R^2$ and Spearman $\rho$ .
- Structural Analysis: Representational Similarity Analysis (RSA) was employed to determine if the geometric structure (relational distances) of the human-rated sensorimotor space could be recovered from the model-predicted space.

Key Contributions

Novel Resource: The authors present the first large-scale, 11-dimensional sensorimotor norming database for 3,000 Mandarin Chinese concepts, covering a broad range of grammatical categories.
Metric Validation: The study introduces and validates the PSE-Sensorimotor metric, extending the PSE framework to include interoception and action effectors, demonstrating its superior predictive power over traditional composites.
Empirical Grounding Data: The paper provides empirical evidence on the recoverability of sensorimotor information from distributional linguistic representations in a non-English language, offering a benchmark for embodied AI research.

Results

Reliability and Validity: The collected norms demonstrated high internal reliability (Cronbach's $\alpha$ ranging from .698 to .949 across dimensions) and strong cross-norm validity with existing Chinese resources.
Descriptive Statistics: Visual strength was the dominant dimension (47.0% of concepts), followed by Interoception (39.1%). Chemosensory dimensions (Gustatory, Olfactory) had the lowest mean ratings but the highest mean strengths for dominant concepts. Embodiment ratings correlated more strongly with action effectors than perceptual modalities.
Lexical Decision Prediction: In Study 2, all sensorimotor composites significantly predicted lexical decision performance. PSE-Sensorimotor and Minkowski-3 emerged as the strongest predictors, outperforming unidimensional measures and simple sums. This suggests that weighting the dominant dimension while retaining contributions from non-dominant dimensions optimally captures the facilitatory effect of sensorimotor information.
Linguistic Recovery: In Study 3, ridge regression models successfully predicted sensorimotor ratings from word embeddings (mean Spearman $\rho$ = .62). Recovery varied by dimension: Visual and Auditory dimensions showed high correspondence, while chemosensory dimensions (Gustatory, Olfactory) showed lower rank-order correlation ( $\rho$ = .39) despite high $R^2$ , attributed to skewed rating distributions and language ineffability.
Structural Recovery: RSA revealed a significant correlation ( $\rho$ = .540) between the relational geometry of the human-rated sensorimotor space and the predicted space, indicating that distributional language patterns partially encode the structure of embodied conceptual knowledge.

Significance
The paper claims that these findings provide a necessary normative resource for advancing embodied cognition research in Mandarin Chinese. By demonstrating that multidimensional sensorimotor composites (specifically PSE-Sensorimotor and Minkowski-3) are robust predictors of lexical processing, the study offers a refined methodological approach for behavioral studies. Furthermore, the partial recoverability of sensorimotor structure from linguistic representations supports the view that language use encodes aspects of embodied conceptual structure, even in the absence of direct physical grounding. This has implications for understanding how large language models might acquire conceptual knowledge and for developing more grounded AI systems. The authors emphasize that these results echo the debate on symbolic grounding, suggesting that while language models lack direct experience, the statistical patterns of language use retain traces of sensorimotor reality.

The Big Idea: How Words "Feel" in the Body

The Project: Building a "Sensory Map" for 3,000 Words

Study 2: The "Speed Test" (Do these feelings help us read faster?)

Study 3: Can a Robot "Feel" Without a Body?

Summary

More like this