indonamegender

Character-level gender classification for Indonesian personal names.

A compact recurrent model that reads a name character by character. It matches fully fine-tuned multilingual encoders at a fraction of the serving cost, and it is the analysis code behind the paper A Cross-Architecture Attention Analysis of Character-Level and Word-Level Models for Gender Classification from Indonesian Names.

Install

pip install indonamegender

Use

from indonamegender import GenderPredictor

p = GenderPredictor()
p.predict("SITI AMINAH")
# {'name': 'SITI AMINAH', 'gender': 'Female', 'confidence': 0.9997, 'model': 'CharBiLSTM'}

The bundled model is CharBiLSTM, the one the paper recommends for names from beyond the training data. It reaches 0.9375 F1 on an independent benchmark and answers in 0.3952 ms on a single CPU thread. No download is needed for it.

Other models

Seven further checkpoints are fetched once from the GitHub release and cached locally, so any of the eight architectures in the grid can be asked for by name.

GenderPredictor("CharBiGRU")
GenderPredictor("WordTransformer")

Test F1 is the mean over five seeds on the 2024 to 2025 partition. CharBiLSTM is bundled, the other seven download on first use. The three pretrained encoders from the paper, IndoBERT, mBERT and XLM-R, are not distributed through this package because their base weights come from Hugging Face. The scripts that fine-tune them are in the repository under experiments/.

Notes

The label is a binary administrative field, male or female. It reflects a recorded administrative label, not how a person identifies, and the model estimates that field rather than a person. Intended for completing missing fields in existing records for aggregate analysis, not for decisions about individuals.

Links

• Source, results, and paper: https://github.com/ericks-rs/indonesian-name-gender
• License: MIT