Abdullah-afify/egy-names-fallback-classifier

egy-names ML Fallback Classifier

A precision-calibrated fallback model for Egyptian names that are not in the egy-names catalog.

egy-names is a book-first engine: 44,626 hand-annotated Egyptian name lemmas with empirical gender, religion, generational-slot, and role labels, derived from 15.88M+ real name records (dataset). The book is always tried first and is always the ground truth.

This model exists only for the names the book has never seen — foreign surnames, rare spellings, brand-new coinages. It infers gender, religion, and role from morphology and character n-grams, and every prediction it returns is explicitly labeled inferred: true so a caller can never confuse a guess with an attested fact.

Why this model is different from a typical classifier

Most name-classification models are tuned to maximize aggregate accuracy. This one is tuned to maximize precision at the moment it chooses to speak, and to abstain (unknown / neutral) rather than guess when it isn't confident enough. That distinction matters for names, where a confidently wrong gender or religion label is worse than no label at all.

Abstention thresholds were not picked by feel — they were measured directly with a held-out precision-at-threshold calibration script and recalibrated after an initial pass under-delivered on its own promise (role="given" was returning only 81.8% precision, religion="christian" only 89.4%, at the original cutoffs):

Class	Abstention threshold	Measured precision at that threshold
gender = male	p ≥ 0.70	93.3%
gender = female	p ≥ 0.70	95.5%
religion = muslim	p ≥ 0.85	96.6%
religion = christian	p ≥ 0.90	95.1%
role = given	p ≥ 0.88	93.0%

Architecture

• Features: character n-grams (TF-IDF) over the normalized surface form, plus hand-built morphological flags (script, common Arabic/English prefixes and suffixes, length, diacritic presence).
• Classifiers: three independent multinomial logistic regression heads — one each for gender, religion, and role — trained on the book's own 44,626 lemmas (excluding low-confidence/fabricated rows) as ground truth.
• Rule pre-checks: a small set of high-precision prefix/suffix rules (e.g. عبد-prefixed compounds → male/Muslim with very high confidence) run before the statistical model and short-circuit it when they fire — the same rule table the book-index detectors use, shared via logic_config.json.
• Nearest-neighbor grounding: every inferred name is also matched to its closest known book entry by Levenshtein distance, so callers can see why the model guessed what it guessed.
• Abstention: below the calibrated threshold, the model returns unknown (religion/role) or neutral (gender) rather than force a label.

Files

• infer_model.json.gz — the exported model weights (TF-IDF vocabulary/IDF, logistic regression coefficients for all three heads) in a small, dependency-free JSON format, runnable from any language without needing scikit-learn at inference time.

Intended use

Fallback-only inference for names absent from the egy-names catalog, inside the egy-names Python SDK's identify()/identify_all() API. Not intended as a standalone general-purpose name classifier outside that context — it was trained and calibrated specifically against Egyptian-Arabic naming patterns.

Training data

Abdullah-afify/egyptian-names — the same 44,626-lemma canonical catalog that powers the egy-names library, filtered to exclude non-personal and low-confidence/fabricated rows before training.

License

MIT — free for academic, commercial, and research use.

@misc{afify2026egynames_fallback,
  author       = {Abdullah Afify},
  title        = {egy-names ML Fallback Classifier},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Abdullah-afify/egy-names-fallback-classifier}}
}