Gender classification from Indonesian personal names

Code, trained checkpoints and result tables for A Cross-Architecture Attention Analysis of Character-Level and Word-Level Models for Gender Classification from Indonesian Names.

The study compares two input representations across four sequence encoders under one fixed configuration, and asks where each model places its attention. It does not propose an architecture. Every model here is a standard encoder used as published.

What the study found

Character-level representation beat word-level representation for all four architectures, by 1.89 to 2.59 F1 points over five matched seeds, with every confidence interval clear of zero. Encoder choice decided far less. Within the character-level group no pairwise difference survived Holm correction, so the paper names no best encoder.

The character-level models match the fine-tuned encoders on the test partition and lead on transfer, at 44.6 to 151.4 times lower single-thread CPU latency. Full results are in results/final/.

Install

pip install -e .

Or with the analysis extras, needed only to regenerate tables and figures.

pip install -e ".[analysis]"

Predict

from indonamegender import GenderPredictor

p = GenderPredictor()
p.predict("GATOTKACA WIRAWAN")
# {'gender': 'Male', 'confidence': 0.999, 'model': 'CharBiLSTM', ...}

The bundled model is the seed-42 CharBiLSTM, which is the one the paper recommends for names from outside the source institution. It has the highest external F1 at 0.9375 and the lowest CPU latency at 0.3952 milliseconds, and it reproduces the stored prediction on all 15,923 evaluation names exactly. It trails CharBiGRU internally by 0.04 points, a difference the paper reports as not separable.

The other seven are under models/ in a clone. From a pip install they are fetched once from the v1.0.0 release and cached under ~/.cache/indonamegender/, so any architecture in the grid can be asked for by name.

GenderPredictor("CharBiGRU")        # bundled? no, downloaded once
GenderPredictor("WordTransformer")  # same

Layout

src/indonamegender/   inference package, with the bundled model
experiments/          the two training entry points
analysis/             statistics, figures and audits over the artifacts
models/               eight seed-42 checkpoints, one per architecture
tokenizers/           character vocabulary, and the hashed word vocabulary
results/final/        every reported number, as CSV
results/figures/      the 24 manuscript figures, PNG and PDF at 600 dpi
configs/              the machine and the training settings the run used
demo/                 local web demo
docs/                 data access, protocol, reproducibility

Reproducing

Training needs the corpus, which is not redistributable. docs/DATA.md says what the inputs are and what a substitute has to satisfy.

python experiments/train_grid.py          # eight architectures, five seeds
python experiments/train_imbalance.py     # three imbalance strategies, 120 runs

Eleven of the twenty-two analysis scripts run against a fresh clone, because they read results/final/ and nothing else. Each was executed from a clean checkout rather than assumed to work.

python analysis/paired_comparison.py        # character against word, paired
python analysis/threshold_free_metrics.py   # AUC and Brier
python analysis/imbalance_protocol.py       # the 120-run robustness check
python analysis/architecture_comparison.py  # within-level encoder comparisons
python analysis/external_paired.py          # the external benchmark, paired
python analysis/audit_personal_data.py      # the privacy scan, over this tree

The remaining eleven stop at their first read. Nine need the corpus, and two need the training layout rather than the released one. They ship because they are the code that produced the reported numbers, not because they can be rerun here, and docs/REPRODUCIBILITY.md says which is which.

docs/PROTOCOL.md gives the split and the training configuration. docs/REPRODUCIBILITY.md reports what is exact and what is not, measured rather than claimed.

Privacy

The corpus is a university admissions registry. No file here contains a registration.

The word vocabulary ships hashed. Every entry except the two reserved indices is blake2s(salt + token), which keeps the embedding index intact while making the 24,947-entry vocabulary unreadable, so a released checkpoint stays usable without carrying the names it was fitted on. Eighteen per-name tables under results/final/ have their name column replaced by a row identifier, the token count and the three-character ending. analysis/audit_personal_data.py scans the tree, and the build fails rather than warns. Illustrative names in the figures, the demo and the tests are drawn from Javanese shadow theatre and appear nowhere in the corpus.

Citation

See CITATION.cff. Licensed MIT.