language:
- bm
task_categories:
- text-generation
pretty_name: Bambara Text Normalization (Bamadaba)
size_categories:
- 1K<n<10K
tags:
- bambara
- text-normalization
bm-text-normalization
Bambara (Bamanankan) orthographic normalisation: map a non-standard spelling to its
standard form. 4,877 short phrase-level pairs in a single config, bamadaba.
Load
from datasets import load_dataset
train = load_dataset("djelia/bm-text-normalization", "bamadaba", split="train")
dev = load_dataset("djelia/bm-text-normalization", "bamadaba", split="dev")
test = load_dataset("djelia/bm-text-normalization", "bamadaba", split="test")
# rows that are already correct — a model must not "fix" these
already_clean = train.filter(lambda row: row["noise_level"] == 0)
The validation split is named dev; split="validation" will fail.
Splits
| Split | Rows |
|---|---|
train |
3,901 |
dev |
488 |
test |
488 |
Fields
| Field | Description |
|---|---|
source_text |
Non-standard spelling |
target_text |
Standard-orthography form |
noise_level |
Character-level edit distance between the two, integer 0-25 |
Example rows:
source_text |
target_text |
noise_level |
|---|---|---|
a bolira coyi ! |
a bolila coyi ! |
1 |
a be lajailen |
a bɛɛ lajɛlen |
4 |
Notes
419 rows are identity pairs at noise_level 0 — text already correct, to be left alone.
The noise is a small closed inventory: Bambara characters replaced by what a French keyboard
produces (ɛ→e, ɔ→o, ɲ→gn), liquid confusion (r/l), plus deleted and inserted
spaces and dropped apostrophes (k'a written ka). Word boundaries move in both directions,
so a model has to join and split, not only map characters.
Sentences are short — median 19 characters. For running text such as ASR transcripts, see
djelia/bm-text-normalization-benchmark.