Research Question: Do multilingual embedding models encode culturally-specific semantic relationships between abstract moral concepts?
Answer: Yes. The geometric relationships between moral concepts vary significantly across languages, suggesting that embedding models absorb cultural values from training data.
The relationship between justice, mercy, and punishment differs by 21% across languages: This difference is statistically significant (F=48.0, p<0.0001).
| Language | Ratio* | Interpretation |
|---|---|---|
| 🇯🇵 Japanese | 1.279 | Justice and mercy closely related |
| 🇬🇧 English | 1.290 | Balanced relationship |
| 🇨🇳 Chinese | 1.308 | Moderate separation |
| 🇸🇦 Arabic | 1.476 | Greater conceptual distance |
| 🇮🇳 Hindi | 1.549 | Justice and mercy are distinct concepts |
*Ratio = distance(justice→mercy) / distance(justice→punishment)
Statistical Validation:
- ANOVA: F = 48.0, p < 0.0001
- Effect Size: 21% variation across languages
- Conclusion: Language is a statistically significant predictor of concept geometry
Methodology
Model: paraphrase-multilingual-mpnet-base-v2 (768-dimensional embeddings)
Languages: English, Hindi, Japanese, Arabic, Chinese
Concepts: 10 abstract moral terms (justice, mercy, duty, honor, forgiveness, punishment, law, freedom, loyalty, sacrifice)
Metric: Cosine distance = 1 - cosine_similarity
Statistical Test: One-way ANOVA with bootstrap resampling (n=20 per language)
Visualization: Interactive Plotly HTML charts
- Japanese embeddings show integrated moral concepts (balance, harmony)
- Hindi embeddings show distinct moral categories (dharma, karma, daya as separate)
- This aligns with cultural scholarship on Eastern vs Western moral philosophy
- Multilingual models are not culturally neutral
- Training data cultural context is encoded in embedding geometry
- Implications for AI fairness, cross-cultural NLP, moral reasoning systems
Shows the 21% variation in concept relationships across languages
Visualizes how moral concepts cluster in Hindi embeddings
Visualizes how moral concepts cluster in Japanese embeddings
- Click the badge above
- Click Runtime → Run all
- Results appear in ~5 minutes
# Clone repository
git clone https://github.com/SShreeya-Das/Embedding-compass.git
cd embedding-compass
# Install dependencies
pip install -r requirements.txt
# Run analysis
python analysis.py # (if you created this file)
# OR open the notebook in Jupyter