SpellMtei

Artifacts for SpellMtei, a spelling corrector for Manipuri (Meetei) in the Meetei Mayek script, trained entirely from synthetic supervision: clean text is corrupted by an error model grounded in the script's structure and in Meetei phonology, and models learn to reconstruct it.

Anonymised for peer review. De-anonymised on acceptance.

Contents

path	what
error_profile/profile.py	the declarative Meetei Mayek error profile (Unicode ranges, combining marks, confusable groups, rewrite rules, per-operation weights). base.py is the framework it plugs into.
eval/test.txt	frozen clean test references (1,012 sentences)
eval/test.noised.tsv	frozen test set: corrupted<TAB>gold, realised character error rate 6.26%
model/manipuri-mm/m1_byt5_small/best/	M1 — google/byt5-small fine-tuned as a denoiser
model/manipuri-mm/m2_gector/	M2 — from-scratch character edit tagger (GECToR-style, non-autoregressive); tags.json / charvocab.json sit next to best/
model/manipuri-mm/m3_kenlm/	M3 — statistical noisy channel: character 6-gram LM (charlm.json) + word-frequency table (wordfreq.json) + tuned params.json
pipeline/	the code: noise engine, datasets, metrics, per-model train/infer, and the evaluation pipeline

The training and validation splits are not included here.

Results (frozen synthetic test set, input CER 6.26%)

System	P	R	F0.5	chrF	CER%	FP-clean%
Uncorrected input	–	–	–	85.0	6.26	0.0
M1 ByT5	35.4	41.3	36.5	91.4	2.86	60.0
M2 char edit tagger	24.7	33.8	26.1	87.8	3.78	86.1
M3 noisy channel	6.4	9.7	6.9	76.9	8.83	97.4

M1 removes 54% of the character error and M2 40%; the statistical channel does not beat leaving the text alone, and all three over-correct already-correct input.

Usage

# drop model/ and pipeline/ into a checkout, then
cd pipeline
python -m pipeline.evaluate all --lang manipuri-mm --split test

License

To be finalised. The clean corpus is public Meetei Mayek web text; provenance and licence will be documented with the camera-ready release.