Copy_Dakshina_Google_research_dataset

This repository is a structured, processed version of the Dakshina Dataset, originally released by Google Research. It has been reorganized into a unified Hugging Face format to support NLP research in South Asian languages, specifically focusing on sentence-level and word-level transliteration tasks.

Dataset Overview

The original Dakshina dataset is a collection of text in both Latin and native scripts for 12 South Asian languages. This Hugging Face dataset specifically extracts and enriches the parallel sentence data (Romanized) and the word-level transliteration mappings (Lexicons) from the original release.

• Languages: 12 South Asian languages.
• Total Rows: 119,758 parallel sentences.
• Format: Parquet / Hugging Face DatasetDict with 12 language subsets.

Included Languages

Language	Code	Script	Sub-Dataset Size
Bangla	bn	Bengali	~10k rows
Gujarati	gu	Gujarati	~10k rows
Hindi	hi	Devanagari	~10k rows
Kannada	kn	Kannada	~10k rows
Malayalam	ml	Malayalam	~10k rows
Marathi	mr	Devanagari	~10k rows
Punjabi	pa	Gurmukhi	~10k rows
Sindhi	sd	Perso-Arabic	~10k rows
Sinhala	si	Sinhala	~10k rows
Tamil	ta	Tamil	~10k rows
Telugu	te	Telugu	~10k rows
Urdu	ur	Perso-Arabic	~9.76k rows

(Note: The Urdu subset contains slightly fewer rows because the original authors manually excluded strings sampled from the Urdu Wikipedia that were determined to be in other languages, such as Arabic or Punjabi).

Dataset Structure

Each language subset contains the following features (columns):

Feature	Type	Description
native	string	The original Wikipedia sentence written in the native script.
romanized	string	The manually annotated, phonetic Latin-script romanization (transliteration) of the native sentence.
native_word_count	int64	The number of whitespace-delimited tokens in the native script sentence.
romanized_word_count	int64	The number of whitespace-delimited tokens in the romanized sentence.
lexicon_mapping	string (JSON)	A JSON string mapping words found in the native sentence to a list of known romanizations, derived from the original dataset's sampled training lexicons.
token_alignment	list[dict]	A sequence of dictionaries representing the precise token-by-token alignment between the native and romanized strings (derived from the *.aligned.cased_nopunct.tsv files).

How This Version Was Generated

This Hugging Face dataset was generated by processing the raw dakshina_dataset_v1.0.tar release via a Python script.

1. Sentence Extraction: Parallel sentences were extracted from the $L/romanized/$L.romanized.rejoined.tsv files, which contain unsegmented native and romanized strings.
2. Feature Engineering: Word counts (native_word_count, romanized_word_count) were calculated dynamically.
3. Lexicon Integration: The training lexicons ($L.translit.sampled.train.tsv) were parsed. For every word in a native sentence, the script searched the lexicon and embedded all attested romanizations into the lexicon_mapping column as a serialized JSON string.
4. Alignment Parsing: Token alignments were parsed from the .aligned.cased_nopunct.tsv files and stored as list objects.

Original Data Collection Details (from Authors)

• Source: All native script strings were sourced from Wikipedia text, filtered to consist primarily of strings in the appropriate Unicode codeblock.
• Romanization: 10,000 strings from each language's validation set were randomly chosen to be romanized by native speaker annotators.
• Lexicons: Lexicon words were sampled from words occurring more than once in the Wikipedia training sets. Most received romanizations from more than one annotator.
• Validation: The authors validated the manual romanizations via a round-trip transcription process, maintaining a character error rate between 3.5 and 8.5 percent.

Citation

If you use or discuss this dataset in your work, please cite the original paper:

@inproceedings{roark-etal-2020-processing,
    title = "Processing {South} {Asian} Languages Written in the {Latin} Script:
    the {Dakshina} Dataset",
    author = "Roark, Brian and
      Wolf-Sonkin, Lawrence and
      Kirov, Christo and
      Mielke, Sabrina J. and
      Johny, Cibu and
      Demir{\c{s}}ahin, I{\c{s}}in and
      Hall, Keith",
    booktitle = "Proceedings of The 12th Language Resources and Evaluation Conference (LREC)",
    year = "2020",
    url = "https://www.aclweb.org/anthology/2020.lrec-1.294",
    pages = "2413--2423"
}

License: The dataset is licensed under CC BY-SA 4.0.