Tencent AI Lab - Shanghai Jiao Tong University Low-Resource Translation System for the WMT22 Translation Task
27415Authors
nan
Abstract
AbstractThis paper describes Tencent AI Lab - Shanghai Jiao Tong University (TAL-SJTU) Low-Resource Translation systems for the WMT22 shared task. We participate in the general translation task on English-Livonian.Our system is based on M2M100 with novel techniques that adapt it to the target language pair.(1) Cross-model word embedding alignment: inspired by cross-lingual word embedding alignment, we successfully transfer a pre-trained word embedding to M2M100, enabling it to support Livonian.(2) Gradual adaptation strategy: we exploit Estonian and Latvian as auxiliary languages for many-to-many translation training and then adapt to English-Livonian.(3) Data augmentation: to enlarge the parallel data for English-Livonian, we construct pseudo-parallel data with Estonian and Latvian as pivot languages.(4) Fine-tuning: to make the most of all available data, we fine-tune the model with the validation set and online back-translation, further boosting the performance. In model evaluation: (1) We find that previous work underestimated the translation performance of Livonian due to inconsistent Unicode normalization, which may cause a discrepancy of up to 14.9 BLEU score.(2) In addition to the standard validation set, we also employ round-trip BLEU to evaluate the models, which we find more appropriate for this task. Finally, our unconstrained system achieves BLEU scores of 17.0 and 30.4 for English to/from Livonian.
Keywords
Full Content
Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1 - 45 December 7-8, 2022 ©2022 Association for Computational Linguistics Findings of the 2022 Conference on Machine Translation (WMT22) Tom Kocmi Microsoft Rachel Bawden Inria, Paris Ondˇrej Bojar Charles University Anton Dvorkovich Neurodub Christian Federmann Microsoft Mark Fishel University of Tartu Thamme Gowda Microsoft Yvette Graham Trinity College Dublin Roman Grundkiewicz Microsoft Barry Haddow University of Edinburgh Rebecca Knowles NRC Philipp Koehn Johns Hopkins University Christof Monz University of Amsterdam Makoto Morishita NTT Masaaki Nagata NTT Toshiaki Nakazawa University of Tokyo Michal Novák Charles University Martin Popel Charles University Maja Popovi´c Dublin City University Mariya Shmatova Neurodub Abstract This paper presents the results of the General Machine Translation Task organised as part of the Conference on Machine Translation (WMT) 2022. In the general MT task, participants were asked to build machine translation systems for any of 11 language pairs, to be evaluated on test sets consisting of four different domains. We evaluate system outputs with human annotators using two different techniques: reference-based direct assessment and (DA) and a combination of DA and scalar quality metric (DA+SQM). 1 Introduction The Seventh Conference on Machine Translation (WMT22)1 was held online with EMNLP 2022 and hosted a number of shared tasks on various aspects of machine translation. This conference built on 15 previous editions of WMT as workshops and conferences (Callison-Burch et al., 2007, 2008, 2009, 2010, 2011, 2012; Bojar et al., 2013, 2014, 2015, 2016, 2017, 2018; Barrault et al., 2019, 2020; Akhbardeh et al., 2021). For more than a decade, the machine translation (MT) community has focused on the news domain, which has many desirable features for MT evalu- ation, such as sufficiently long and grammatically 1http://www.statmt.org/wmt22/ correct sentences that are easy for both profession- als to translate (to produce references) and for hu- man raters to evaluate without specific in-domain knowledge. However, with recent advances in MT and potential overfitting on the news domain (with methods such as fine-tuning on past WMT testsets), we decided to open a fresh research direction of testing the “General Machine Translation” capabil- ities. How to test general MT capabilities is a research question in itself. Countless phenomena could be evaluated, the most important being: • various domains (news, medicine, IT, patents, legal, social, gaming, etc.) • style of text (formal or spoken language, fic- tion, technical reports, etc.) • noisy or robust user-generated content (gram- matical errors, code-switching, abbreviations, etc.) Evaluating all possible phenomena is near im- possible and creates many unforeseen problems. Therefore, we decided to simplify the problem and start with an evaluation of different domains. We select the following four domains: news, e- commerce, social, and conversational, chosen to represent various topics with different content 1 styles. Additionally, these domains are understand- able for humans without special in-domain knowl- edge, thus not requiring specialized translators or human raters for evaluation. Another significant change for this year is the redesign of our human evaluation procedure for English→X and non-English language pairs. …
Content truncated to 500 words
Labeled Text
AbstractThis paper describes Tencent AI Lab - Shanghai Jiao Tong University (TAL-SJTU) Low-Resource Translation systems for the WMT22 shared task. We participate in the general translation task on English-Livonian.Our system is based on M2M100 with novel techniques that adapt it to the target language pair.(1) Cross-model word embedding alignment: inspired by cross-lingual word embedding alignment, we successfully transfer a pre-trained word embedding to M2M100, enabling it to support Livonian.(2) Gradual adaptation strategy: we exploit Estonian and Latvian as auxiliary languages for many-to-many translation training and then adapt to English-Livonian.(3) Data augmentation: to enlarge the parallel data for English-Livonian, we construct pseudo-parallel data with Estonian and Latvian as pivot languages.(4) Fine-tuning: to make the most of all available data, we fine-tune the model with the validation set and online back-translation, further boosting the performance. In model evaluation: (1) We find that previous work underestimated the translation performance of Livonian due to inconsistent Unicode normalization, which may cause a discrepancy of up to 14.9 BLEU score.(2) In addition to the standard validation set, we also employ round-trip BLEU to evaluate the models, which we find more appropriate for this task. Finally, our unconstrained system achieves BLEU scores of 17.0 and 30.4 for English to/from Livonian.
Publication Details
- Paper ID:
- 27415
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan