No Domain Left behind

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 423 - 427 December 7-8, 2022 ©2022 Association for Computational Linguistics Abstract We participated in the WMT General MT task and focus on four high resource language pairs: English to Chinese, Chinese to English, English to Japanese and Japanese to English). The submitted systems (LanguageX) focus on data cleaning, data selection, data mixing and TM-augmented NMT. Rules and multilingual language model are used for data filtering and data selection. In the automatic evaluation, our best submitted English to Chinese system achieved 54.3 BLEU score and 63.8 COMET score, which is the highest among all the submissions. 1 Introduction Training neural machine translation models for a specific domain is a well-studied task. However, maximizing the performance of a single NMT model for multiple domains remains difficult. As a former translator and a current machine translation engineer, I always dream about building a versatile machine translation system – no domain left behind. Our neural machine translation system is developed using big transformer (Vaswani et al., 2017) architecture and the toolkit I used is fairseq (Ott et al., 2020). Rules, multilingual language model and faiss (Johnson et al., 2021) are used to align, clean and select parallel data. The following techniques are used in model training: a. Data mixing is used to mix general domain corpus with specific domain corpus; b. Back translation (Sennrich et al., 2016) is not applied because it is time-consuming. Instead, Neural Machine Translation with Monolingual Translation Memory (Cai et al., 1http://mteval.cipsc.org.cn:81/agreement/description 2021) is used to fully utilize the monolingual corpus. 2 Data Filtering and Selection The Chinese-English parallel data is mainly from CCMT Corpus 1 , inhouse domain data from translation projects, as well as parallel data aligned from multilingual websites and e-books. The monolingual data for multiple domains is collected from the internet and e-books. WMT newstest2021 is used to evaluate the model’s general domain performance. Multiple domain- specific test sets are created to evaluate the model’s specific domain performance. Each domain has a test set of 1,000 sentences. In order to build a versatile machine translation system, a total of 15 domains are covered in preparing the parallel and monolingual corpus. The primary domains and subdomains are listed as follows: Literature Web novel Famous literary work Literature/Poetry Idioms/maxims/sayings Slang Conversation Names (personal, company) Symbols / Abbreviations / Acronyms Art, History and Philosophy Arts/crafts/painting Cooking/culinary/gastronomy Folklore History Philosophy Graphic arts/photo/imaging Music Religion No Domain Left Behind Hui Zeng LanguageX AI Lab felix_zeng_ai@aliyun.com 423 Social Science, Sociology, Ethics, etc. Economy, Finance and Business Business/commerce Accounting Finance (general) Investment / Securities Insurance Economics Real Estate Fashion and Marketing Advertising / Public Relations Marketing / Market Research Cosmetics / Beauty Fashion Textiles / Clothing / Fashion Clothing/textiles Politics and National Defense Government/politics International org/Dev/coop Military / Defense Law Law (general) Law: Contract(s) Law: patents/trademarks/copyrights Law: Taxation & Customs Computers and IT (Information Technology) Computers (general) Computers: Systems, Networks Computers: Hardware IT (Information Technology) Telecommunications Internet, e-Commerce SAP System Applications and …

Content truncated to 500 words