InterpretCognates

Interactive web platform, 135-language embedding norm suite, and deterministic statistical verification pipeline for probing how multilingual neural translation networks represent universal conceptual structure across languages.

• Live Interactive Companion: https://kylemath.github.io/InterpretCognates/ (22 synchronized 2D/3D Plotly charts, phylogenetic trees, colexification graphs, and layerwise emergence explorers).
• Preprint (arXiv v2): Universal Conceptual Structure in Neural Translation: Probing NLLB-200's Multilingual Geometry (paper/main.pdf; see paper/REPRODUCING.md to build from source).
• Permanent Archive (Zenodo DOI): 10.5281/zenodo.23898624
• Journal Submission: Submitted to Behavior Research Methods (Manuscript ID: BR-Org-26-1447).
• Companion Software Paper (JOSS): joss/paper.md.
• Screenshot — App overview.

1. Additive Language-Shift + Conceptual-Manifold Geometry: Two-way Concept $\times$ Language variance partitioning and eigenvalue Participation Ratio (PR) dimensionality analysis show that NLLB-200 separates low-dimensional per-language shifts ($\text{PR} = 13.3$, capturing $45.7%$ of native bare-word variance vs. $13.7%$ in carrier frames) from an orthogonal, nearly $3\times$ higher-dimensional conceptual manifold ($\text{PR} = 38.7$). Subtracting per-language centroids increases the Conceptual Store Metric (CSM) by $2.14\times$ on native bare words ($1.15 \to 2.46$, vs. a $0.99$ permutation null; $2.27\times$ across disjoint scripts) and $1.19\times$ in carrier frames ($2.25 \to 2.69$).
2. Layerwise Emergence & POS-Selective Terminal Unpacking: Tracing representations across Layers $0\dots 11$ reveals a sigmoid phase transition (inflection at Layer 6, peaking at Layer 10 with $4.35\times$ mutual $k$-NN alignment gain), followed by POS-selective cone expansion at Layer 11 driven by pronouns and verbs preparing for decoder cross-attention.
3. Semantic Dominance & Zipfian Subword Mediation: Semantic category explains $\eta^2 = 25.4%$ of cross-lingual convergence variance ($p = 2.21 \times 10^{-4}$)—over $20\times$ orthographic or phonological form ($R^2 \le 0.012$)—while Swadesh noun convergence over non-Swadesh nouns ($d = 0.54\text{--}0.69$) is mechanistically mediated by Zipfian subword compactness ($p = 7.43 \times 10^{-7}$).
4. Phylogenetic (ASJP) & Lexical Typology ($\text{CLICS}^3$) Benchmarks: Embedding distances track ASJP phonetic distances within subfamilies ($\rho_s = 0.65$, Jäger's LDND $\rho_s = 0.67$) and recurrent cross-family $\text{CLICS}^3$ colexifications ($d = 1.43, p = 1.55 \times 10^{-33}$).
5. Psycholinguistic Parallels: Basic color terms recover the Berlin & Kay evolutionary hierarchy ($\rho_s = 0.91\text{--}0.93$), and high- vs. lower-resource training tiers exhibit asymmetric conceptual convergence ($p \le 3.63 \times 10^{-9}$) analogous to Kroll & Stewart's (1994) Revised Hierarchical Model.
6. Zero-Hardcoded-Numbers Reproducibility: All 321 reported statistics in paper/main.pdf are programmatically generated from docs/data/*.json into paper/output/stats.tex via paper/scripts/generate_stats.py.

If you use InterpretCognates, its interactive web explorer, or its 135-language geometric norms, please cite:

@article{mathewson2026universal,
  title={Universal Conceptual Structure in Neural Translation: Probing {NLLB-200}'s Multilingual Geometry},
  author={Mathewson, Kyle E.},
  journal={arXiv preprint arXiv:2603.02258},
  year={2026},
  url={https://arxiv.org/abs/2603.02258},
  note={Submitted to Behavior Research Methods (BR-Org-26-1447)}
}

This toolkit uses Meta NLLB-200 (open weights) through Hugging Face Transformers:

• Practical default in code: facebook/nllb-200-distilled-600M (fast local iteration).
• Midpoint quality option: facebook/nllb-200-distilled-1.3B.
• Highest quality option: facebook/nllb-200-3.3B (requires larger GPU/CPU RAM).

Set the model with:

export NLLB_MODEL=facebook/nllb-200-distilled-1.3B

From repository root:

cd backend
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload

Open http://127.0.0.1:8000.

cd paper
python3 scripts/generate_stats.py
python3 scripts/generate_figures.py
bash build.sh
bash build_arxiv.sh

The interactive research companion runs entirely client-side from pre-computed JSON data in docs/data/:

python3 build_static.py          # outputs to docs/
python3 build_static.py dist     # or specify a custom output directory

Released under the MIT License.