README.md

metadata

license: mit
task_categories:
  - text2text-generation
language:
  - ml
size_categories:
  - 1K<n<10K

Romanized to Native Malayalam Script Transliteration

Shared-Task code submitted to IndoNLP Workshop colocated with COLING 2025, in Abu Dhabi, UAE.

The training notebook shows the code.

Training

Data derived from:

1. Dakshina
   1. Romanized - 186 thousand word transliteration pairs
   2. [Lexicons] (https://huggingface.co/datasets/vrclc/dakshina-lexicons-ml) - 58.4 thousand lexicon entries
2. Aksharantar - 4.3 million word transliteration pairs

RESULT

We present the evaluation scores averaged over the two test datsets shared by the task organizers.

Dataset	CER (%)	WER (%)	BLEU (%)
Test Set-1	7.4	34.5	32.7
Test Set-2	22.7	66.9	7.5

Test Set-1

Test Set-2