PROMT Systems for WMT22 General Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 342 - 345 December 7-8, 2022 ©2022 Association for Computational Linguistics Abstract This paper describes the PROMT submissions for the WMT22 Shared General Translation Task. This year we participated in four directions of the Shared Translation Task: English to Russian, English to German and back, and Ukrainian to English. All our models are trained with the MarianNMT toolkit using the transformer-big configuration. We use BPE for text encoding, all of our models are unconstrained. We achieve competitive results according to automatic metrics in all directions. 1 Introduction The WMT Shared General Translation Task is an annual event where different companies and researchers build and test their systems on the test sets provided by the organizers. This year the Task has shifted from news to the general domain. We participate in four directions: English to Russian, English to German and back, and Ukrainian to English. We build the transformer-big models for the first time. We also explore new data filtering techniques, data preparation and model training strategies. The rest of the paper is organized as follows: in Section 2 we describe in detail the systems we submitted to the Shared Task. In Section 3 we present and discuss the results. We conclude the paper in Section 4 with discussion for possible future work. 2 Systems overview All of our WMT22 submissions are MarianNMT- trained (Junczys-Dowmunt et al., 2018) transformer-big (Vaswani et al., 2017) systems. We use the OpenNMT toolkit (Klein et al., 2017) version of byte pair encoding (BPE) (Sennrich et al., 2016b) for subword segmentation. Our BPE models are case-insensitive, we use special tokens in the source and target sides to process case (see Molchanov (2019) for details). All of the systems are unconstrained, i.e. we use all data provided by the WMT organizers, all publicly available data and some private data crawled from different web-sources. This year we use the dual conditional cross- entropy (Junczys-Dowmunt, 2018) method for data filtering. We extend the method as proposed by the author and build neural language models for both source and target languages. We also augment our training data with two types of synthetic data: 1) back-translations (Sennrich et al., 2016a) and 2) synthetic data with placeholders as described in Pinnis et al. (2017). The back-translations are obtained using the previous versions of our NMT models which are baseline transformers trained with less data (and without some up-to-date data like the news 2021 corpora from statmt.org). We also tag all our synthetic data with special tokens at the beginning of the source sentences as described in Caswell et al. (2019). All models are trained with guided alignment which is used at translation time to handle named entities and document formatting. We obtain PROMT Systems for WMT22 Shared General Translation Task Alexander Molchanov, Vladislav Kovalenko & Natalia Makhamalkina PROMT LLC 17E Uralskaya str. building 3, 199155, St. Petersburg, Russia First.Last@promt.ru 342 alignments using the fast-align (Dyer et al., 2013) tool. The data statistics …

Content truncated to 500 words