CUNI-Bergamot Submission at WMT22 General Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 280 - 289 December 7-8, 2022 ©2022 Association for Computational Linguistics CUNI Submission in WMT22 General Task Josef Jon Charles University surname@mail.ufal.mff.cuni.cz Abstract We present CUNI-Bergamot submission for WMT22 General translation task. We compete in English → Czech direction. Our submission fur- ther explores block backtranslation techniques. In addition to the previous work, we measure perfor- mance in terms of COMET score and named enti- ties translation accuracy. We evaluate performance of MBR decoding compared to traditional mixed backtranslation training and we show possible syn- ergy when using both of the techniques simultane- ously. The results show that both approaches are effective means of improving translation quality and they yield even better results when combined. 1 Introduction This work focuses on exploring of two methods used in NMT in order to improve translation qual- ity: backtranslation and Minimum Bayes Risk de- coding using neural-based evaluation metric as a utility function. The methods used and related work are presented in the following section. In next section we describe our experimental setting and results. 2 Methods We describe methods we used to build our system in this section. 2.1 Block backtranslation The translation quality of NMT depends heavily on the amount of parallel training data. It has been shown that the authentic bilingual data can be partially supplemented by synthetically paral- lel, machine translated monolingual text (Bojar and Tamchyna, 2011; Sennrich et al., 2016; Xie et al., 2018; Edunov et al., 2018). Often the synthetic and authentic parallel data are mixed in the training dataset, but previous research shows that simply mixing the two types of text does not yield optimal translation quality. We are using block backtrans- lation (block-BT) in similar configuration to Popel et al. (2020). This method creates blocks of par- allel and synthetic data and presents them to the neural network separately, switching between the two types during the training. Since in last year’s WMT, the submission using block-BT by Gebauer et al. (2021) did not find any improvements, pre- sumably due to improperly chosen block size, we decided to verify effectiveness of this method once again. Averaging type Previous work on block-BT shows the importance of averaging the checkpoints to combine information from different blocks of training data in order to obtain good performance. We compare checkpoint averaging with another method of combining older sets of model’s param- eters with the current one – exponential smoothing. After each update u, the current parameters Θu are averaged (with smoothing factor α) with parame- ters after the previous update Θu−1: Θu = αΘu + (1 − α)Θu−1 Previous work by Popel (2018) contains experi- ments with exponential averaging, but only on the level of already saved checkpoints, not online dur- ing the training after each update as for our work. Minimum Bayes Risk Decoding NMT models predict conditional probability distribution over translation hypotheses given a source sentence. To select the most probable translation under the model (mode of the …

Content truncated to 500 words