PROMT Systems for WMT21 Terminology Translation Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 835–841 November 10–11, 2021. ©2021 Association for Computational Linguistics 835 Abstract This paper describes the PROMT submissions for the WMT21 Terminology Translation Task. We participate in two directions: English to French and English to Russian. Our final submissions are MarianNMT-based neural systems. We present two technologies for terminology translation: a modification of the Dinu et al. (2019) soft-constrained approach and our own approach called PROMT Smart Neural Dictionary (SmartND). We achieve good results in both directions. 1 Introduction The currently state-of-the-art approach of neural machine translation (NMT) does not inherently allow for explicit control over the system’s output. That is why terminology translation has always been a problem for NMT systems. There are several approaches to solving this problem. One common paradigm is constrained decoding (Hokamp and Liu, 2017; Anderson et al., 2017; Post and Vilar, 2018), where the terminology matches are presented as hard constraints that the beam search must satisfy. Constrained decoding has its disadvantages: it is computationally expensive and can deteriorate the translation quality (Dinu et al., 2019). Another common approach is the one introduced by (Dinu et al., 2019): the terminological constraints are provided as input to the NMT as additional annotations inline with the source sentence. These can be considered ‘soft’ constraints, as there is no guarantee that the NMT system will indeed produce an output containing them. In this paper we describe two approaches to terminology translation. First, we propose a modification of the (Dinu et al., 2019) approach. Second, we introduce our own technology PROMT Smart Neural Dictionary (SmartND) aimed at handling terminology translation. The paper is organized as follows: in Section 2 we describe the systems built for the Task and the data we used. In Section 3 we describe our technologies for terminology translation. In Section 4 we present and discuss the results. We conclude the paper with discussion for possible future work in Section 5. 2 Systems overview We submitted two single baseline transformer- based (Vaswani et al., 2017) systems trained with the MarianNMT (Junczys-Dowmunt et al., 2018) toolkit: English-Russian and English-French. We use all parallel data allowed by the organizers. The final systems have the same architecture: we use a shared vocabulary of sizes 16k and 32k for the English-French and English-Russian systems respectively. We use the OpenNMT toolkit (Klein et al., 2017) version of byte pair encoding (BPE) (Sennrich et al., 2016b) for subword segmentation. We use the devsets provided by the organizers as our development sets. We build intermediate models to obtain back- translations (Sennrich et al., 2016a) for our final systems. We use iterative back-translation for the English-Russian system. The intermediate models are trained using SentencePiece (Kudo and Richardson, 2018) for subword segmentation as we noticed that SentencePiece-based models are more robust in low and middle-resource PROMT Systems for WMT21 Terminology Translation Task Alexander Molchanov, Vladislav Kovalenko & Fedor Bykov PROMT LLC 17E Uralskaya str. building 3, 199155, St. Petersburg, Russia First.Last@promt.ru 836 conditions. We also …

Content truncated to 500 words