Coursera Corpus Mining and Multistage Fine-Tuning for Improving Lectures Translation

Overview

This repo is for our paper Coursera Corpus Mining and Multistage Fine-Tuning for Improving Lectures Translation.

It contains both the dataset and all source codes in the paper.

Keywords: Japanese-English parallel dataset, educational domain machine translation, lectures translation, multistage fine-tuning

Dataset

	#lines	#docs	Description
Test	2068	50	Human-validated
Dev	555	16	Human-validated
Train	50543	818	Automatic aligned High quality

Table 1: English-Japanese parallel dataset in educational domain.

	#lines	#docs	Description
Test	2009	90	Human-validated
Dev	865	34	Human-validated
Train	40074	997	Automatic aligned High quality

Table 2: English-Chinese parallel dataset in educational domain.

It contains high quality English-Japanese parallel sentences and documents from site Coursera. Please refer our paper for details.

Update: We updated the English-Japanese dataset and it contains more sentences. We added a new English-Chinese dataset.

Source code

Also it contain the source codes described in the paper:

1. Crawling multi-language subtitle documents from Coursera using youtube-dl.
2. Extracting subtitle files of the desired language pair, data normalization and data clean.
3. Using machine translation and sentence embedding combined with DP to extract parallel sentence pairs in comparable document pairs.
4. Multistage fine-tuning techniques to leverage out-of- and in- domain data to train a MT system for lectures domain translation.

Experiment results

	Ja->En	En->Ja
Coursera dataset only	6.2	6.4
Combined with OOD datasets	27.5	18.5

	Zh->En	En->Zh
Coursera dataset only	14.8	14.5
Combined with OOD datasets	29.5	29.1

Table 2: BLEU scores of using only Coursera dataset and combined with ASPEC, TED Talks datasets for Japanese-English and news commentary, TED Talks for Chinese-English with multistage fine-tuning techniques. Please refer our paper for details.

Reference

Please cite our paper if you used our code or dataset:

@inproceedings{song-etal-2020-coursera,
    title = "{C}oursera Corpus Mining and Multistage Fine-Tuning for Improving Lectures Translation",
    author = "Song, Haiyue  and
      Dabre, Raj  and
      Fujita, Atsushi  and
      Kurohashi, Sadao",
    booktitle = "Proceedings of the 12th Language Resources and Evaluation Conference",
    month = may,
    year = "2020",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://www.aclweb.org/anthology/2020.lrec-1.449",
    pages = "3640--3649",
    language = "English",
    ISBN = "979-10-95546-34-4",
}

Contact

If you have any question, please contact song@nlp.ist.i.kyoto-u.ac.jp