Automatic Post-Editing for Vietnamese

Automatic Post-Editing for Vietnamese Thanh Vu1 and Dai Quoc Nguyen2∗ 1Oracle Digital Assistant, Oracle, Australia 2Oracle Labs, Oracle, Australia 1thanh.v.vu@oracle.com; 2dai.nguyen@oracle.com Abstract Automatic post-editing (APE) is an important remedy for reducing errors of raw translated texts that are produced by machine transla- tion (MT) systems or software-aided transla- tion. In this paper, we present a systematic ap- proach to tackle the APE task for Vietnamese. Specifically, we construct the first large-scale dataset of 5M Vietnamese translated and cor- rected sentence pairs. We then apply strong neural MT models to handle the APE task, us- ing our constructed dataset. Experimental re- sults from both automatic and human evalua- tions show the effectiveness of the neural MT models in handling the Vietnamese APE task. 1 Introduction Recent research has placed significant advance- ments for automatic machine translation (Wu et al., 2016; Vaswani et al., 2017; Barrault et al., 2019). The high-quality MT output has been widely adopted by professional translators into their trans- lation workflow to save time and reduce translation errors (Zaretskaya et al., 2016). Translating Chinese novels to Vietnamese is an important task. In the last ten years, there are about 30K Chinese novels describing fiction stories, that are available in Vietnamese with ∼ 80K active readers and ∼ 600K novel chapter views daily from the three most popular Vietnamese websites for reading novels.123 But, translating the Chinese novels to Vietnamese is still challenging. The rea- son is that in fact, readers prefer reading the novels translated using the traditional language style rather than the modern language style used in news ar- ticles (e.g. using “tiểu nữ nhi”little girl instead of ∗Most of the work was done before two authors joined Oracle. 1https://truyencv.com 2https://truyenyy.com 3https://truyen.tangthuvien.vn “cô bé”little girl). Note that current general-purpose MT systems (e.g., Google Translate), trained on modern language style-focused bilingual corpora, cannot satisfy the reader preference. The well-known workflow/guideline used for translating the Chinese novels to Vietnamese con- sists of three steps:4 • In the first step, the Chinese text is converted into Sino-Vietnamese (i.e. Han-Viet)5 text using a specialized software, such as TTV Translator.6 • In the second step, the Sino-Vietnamese text is further smoothed by replacing predefined Sino- Vietnamese phrases by dictionary-based Viet- namese phrases. The core content of the Viet- namese text generated as the output of the second step—namely software-aided translated text— can be generally understood by frequent readers who are familiar with reading the translated text. Note that the translated text does not fully follow the Vietnamese grammar and vocabulary, thus making it hard for new readers (and even fairly often for the frequent readers) to understand de- tails of the text content. • In the final step, the translated text is manually edited and polished following Vietnamese vo- cabulary and grammar. Here, we refer to the text generated as the output of the final step as the human-corrected text that can be accessed easily by readers with different reading levels. Note that the final editing step is very time- consuming due …

Content truncated to 500 words