PhoMT: A High-Quality and Large-Scale Benchmark Dataset for Vietnamese-English Machine Translation

Abstract

We introduce a high-quality and large-scale Vietnamese-English parallel dataset of 3.02M sentence pairs, which is 2.9M pairs larger than the benchmark Vietnamese-English machine translation corpus IWSLT15. We conduct experiments comparing strong neural baselines and well-known automatic translation engines on our dataset and find that in both automatic and human evaluations: the best performance is obtained by fine-tuning the pretrained sequence-to-sequence denoising autoencoder mBART. To our best knowledge, this is the first large-scale Vietnamese-English machine translation study. We hope our publicly available dataset and study can serve as a starting point for future research and applications on Vietnamese-English machine translation. We release our dataset at: https:// github.com/VinAIResearch/PhoMT.

A High-Quality and Large-Scale Dataset for English-Vietnamese Speech Translation

“…https://github.com/jiaaro/pydub3 We use the English-to-Vietnamese translation direction because Google-translating from English to Vietnamese produces better translation than from Vietnamese to English. This is confirmed via BLEU scores in the first two rows in Tables3 and 4from[23] or humanevaluation results for Google Translate in Table2from[22] 4. For checking the timestamp misalignment, we reuse our PyDubbased tool developed for correcting the timestamps of the first and/or last words of sentences in the third phase (Section 2.3).…”

“…To align parallel sentences within a parallel English-Vietnamese document pair, following [22], we first use Google Translate to translate English source sentences into Vietnamese. 3 Then, to produce parallel English-Vietnamese sentence pairs, we use three alignment toolkits of Hunalign [24], Gargantua [25] and Bleualign [26] to perform an intermediate alignment between the Vietnamese Google-translated versions of the English source sentences and the Vietnamese target sentences.…”

“…For ASR, we train the Fairseq's S2T Transformer model [27,28] on our English audio-transcript training data. 5 For MT, we fine-tune the pretrained sequence-to-sequence model mBART [32] that obtains the state-of-the-art performance for English-Vietnamese machine translation [22,33].…”

“…tuning mBART. We also fine-tune mBART on a combination of the 3M-sentence-pair dataset PhoMT [22] and our extended set (here, we remove PhoMT's pairs that appear in our validation and test sets).…”

Enriching Biomedical Knowledge for Low-resource Language Through Translation

“…One of the first notable parallel datasets and En-Vi neural machine translation is ISWLT’15 (Luong and Manning, 2015) with 133K sentence pairs. A few years later, PhoMT (Doan et al, 2021) and VLSP2020 (Ha et al, 2020) released larger parallel datasets, extracted from publicly available resources for the English-Vietnamese translation.…”