Abstract: Evidence from psychology and cognitive neuroscience indicates that the human brain's semantic system contains several specific subsystems, each representing a particular dimension of semantic information. This article reports a large database, the Six Semantic Dimension Database (SSDD), which contains subjective ratings for 17,940 commonly used Chinese words on six major semantic dimensions: vision, motor, socialness, emotion, time, and space. Furthermore, using computational models to learn the mapping relations between subjective ratings and word embeddings, the authors include estimated semantic ratings for 1,427,992 Chinese and 1,515,633 English words in the SSDD. The SSDD will aid studies on natural language processing, text analysis, and semantic representation in the brain.
Introduction
Cognitive neuroscience suggests that our brains do not store word meanings in a single, monolithic "folder." Instead, semantic representation is distributed across specialized subsystems. For example, the word "apple" triggers visual areas related to color and shape, as well as motor areas associated with the act of biting.
While researchers have previously built databases of these "semantic ratings," they typically cover only a few hundred or thousand words. This creates a significant bottleneck: analyzing the neural response to natural speech or a full novel often fails because most words in the text are missing from the database. The Six Semantic Dimension Database (SSDD) addresses this "sparsity problem" by providing a high-coverage map of the human semantic space for both Chinese and English.
Methodology
The authors employed a two-stage approach to build the SSDD:
- Human Subjective Ratings: 80 native Chinese speakers rated 17,940 high-frequency words across six dimensions: Vision, Motor, Socialness, Emotion, Time, and Space. Most used a 7-point scale, while Emotion used a 13-point valence scale (from very negative to very positive).
- Computational Extension: To scale the resource, the team used word embeddings from models like Word2vec and MacBERT. They trained Ridge Regression models to learn the mapping between these high-dimensional vectors and the human-provided ratings.
- Cross-Lingual Projection: To generate English ratings, they aligned English word embeddings with the Chinese embedding space using a bilingual dictionary and applied their trained prediction models.
Data Distribution and Coverage
The SSDD Subjective Set includes 17,940 commonly used Chinese words, while the Extended Set provides estimated ratings for over 1.4 million Chinese words and 1.5 million English words. The distribution of these ratings highlights the specific nature of semantic information across the vocabulary.
As shown in the histograms, many words cluster at the lower end of specific dimensions, reflecting that most words do not strongly evoke specific sensory or motor imagery. The researchers derived a seventh dimension, "Emotion_abs+1," by taking the absolute value of the Emotion rating. This effectively measures emotional intensity or arousal, aligning it with the 7-point scale used for other dimensions.
Key Insights and Results
- Independence of Dimensions: Most of the six dimensions are relatively independent. However, Vision showed moderate correlations with Motor ($r=0.49$) and Space ($r=0.40$), suggesting that our visual system is deeply intertwined with how we perceive movement and spatial relationships.
- Cross-Language Universality: One of the most striking findings was the high cross-language validity. Human semantic ratings for Chinese words were strongly correlated ($r > 0.7$) with their English translations, suggesting these dimensions capture fundamental aspects of human experience that transcend specific languages.
- Model Performance: Transformer-based models like MacBERT outperformed Word2vec in predicting ratings, likely because they capture more nuanced contextual information. The computational estimates for the 1.5 million English words showed external validity scores very close to those of the human-rated Chinese set.
Conclusion
The SSDD provides a massive, high-quality resource for researchers in natural language processing, text analysis, and cognitive neuroscience. By bridging the gap between small-scale cognitive studies and the "big data" requirements of modern AI, it enables a more comprehensive analysis of how the human brain represents meaning. Future applications may include using this database to identify "semantic signatures" in the language produced by patients with neurological or psychiatric conditions.
Summary updated