Datasets:

jdt-pxl
/

lancorp-augmented

license: cc-by-nc-4.0
language:
  - zh
  - en
tags:
  - hokkien
  - philippine-hokkien
  - southern-min
  - low-resource
  - speech-translation
  - parallel-corpus
  - code-switching
pretty_name: LanCorp Augmented (Philippine Hokkien–English Parallel Corpus)
size_categories:
  - 10K<n<100K
task_categories:
  - automatic-speech-recognition
  - translation

LanCorp Augmented — Philippine Hokkien–English Parallel Dataset

Dataset Summary

This dataset is an augmented version of the Lannang Corpus (LanCorp), a sociolinguistic corpus of Philippine Hokkien (Lán-nâng-uē) speech collected from Filipino-Chinese communities in metropolitan Manila between 2016 and 2020. The augmentation adds clause-level English translations contributed by bilingual annotators, along with word count and duration metadata, and a separate translator provenance file mapping utterances to their respective annotators.

To the best of the author's knowledge, this constitutes the first parallel Philippine Hokkien–English dataset in the public domain.

This dataset was constructed as part of an undergraduate thesis investigating cross-dialect transfer learning for Philippine Hokkien–English speech-to-speech translation at the University of the Philippines - Cebu.

Dataset Structure

The dataset consists of two files:

`utterances.csv`

The augmented LanCorp CSV. Contains all original LanCorp columns plus the following additions:

Column	Type	Description
`english.trans`	`string`	Clause-level English translation contributed by bilingual annotators.
`word.count`	`int`	Word count of the English translation.
`duration_seconds`	`float`	Duration of the associated source audio segment in seconds.

All original LanCorp columns (utterance ID, audio path, transcript, language annotation, part-of-speech tags, social metadata, etc.) are preserved as-is.

`file_translators_anonymized.csv`

A separate file mapping source audio files to translator identifiers, indicating which annotator was responsible for each file's translations. Translator IDs are anonymized.

Column	Type	Description
`filename`	`string`	Source audio filename, matches the `filename` column in the main CSV.
`translator_id`	`string`	Anonymized identifier for the annotator who contributed the translations for this file.
`completion_status`	`string`	Translation completion status for the file; either `partial` (some utterances translated) or `complete` (all utterances translated).

Corpus Filtering

The translated subset used for the associated study was derived by applying the following filters to the full augmented corpus:

Rows without an associated audio file were removed.
Language annotations were normalized to lowercase; only utterances annotated as Hokkien (h) or Lán-lâng-uē (l) were retained.
Utterances containing non-lexical vocalizations (e.g., laughter, fillers) were removed via annotation markers and rule-based filtering.
Rows with duplicate utterance IDs were dropped; duplicate surface forms across different speakers were retained.
Utterances without a confirmed English translation were discarded.

After filtering, 13,404 utterances with confirmed English translations remained, totaling approximately 18.43 hours of audio.

Source Data

This dataset is derived from the Lannang Corpus (LanCorp):

Wilkinson Daniel Wong Gonzales. “Advancing Sino-Philippine linguistics and sociolinguistics using the Lannang Corpus (LanCorp) A multilingual, POS-tagged, and audio-textual databank”. In: International Journal of Corpus Linguistics 29.2 (2024), pp. 213–257

LanCorp is released under CC BY-NC 4.0. This dataset inherits the same license.

Associated Work

This dataset was constructed for use in:

PENDING

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

This dataset is derived from LanCorp (CC BY-NC 4.0) and inherits the same license terms. You are free to share and adapt this dataset for non-commercial purposes, provided appropriate credit is given and the same license is applied to derivative works.

Citation

If you use this dataset, please cite the associated paper and the original LanCorp:

% Associated Paper
% -- PENDING --

% Original LanCorp
@article{gonzales2024advancing,
  title={Advancing Sino-Philippine linguistics and sociolinguistics using the Lannang Corpus (LanCorp) A multilingual, POS-tagged, and audio-textual databank},
  author={Gonzales, Wilkinson Daniel Wong},
  journal={International Journal of Corpus Linguistics},
  volume={29},
  number={2},
  pages={213--257},
  year={2024},
  publisher={John Benjamins Publishing Company Amsterdam/Philadelphia}
}