KYB General Machine Translation Systems for WMT22

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 290 - 294 December 7-8, 2022 ©2022 Association for Computational Linguistics Abstract We here describe our neural machine translation system for the general machine translation shared task in WMT 2022. Our systems are based on the Transformer (Vaswani et al., 2017) with base settings. We explore the high-efficiency model training strategies, aimed to train a model with high-accuracy by using a small model and a reasonable amount of data. We performed fine-tuning and ensembling with N-best ranking in English to/from Japanese directions. We found that fine-tuning by filtered JParaCrawl data set leads to better translations for both directions in English to/from Japanese models. In the English to Japanese direction model, ensembling and N-best ranking of 10 different checkpoints improved translations. By comparing with another online translation service, we found that our model achieved a great translation quality. 1 Introduction We participated in the Japanese to/from English translation for the general machine translation shared task of WMT 2022. Japanese →English is one of the challenging language pairs for machine translation since their differences are large in both vocabulary and grammatical structure. Recent advances in neural machine translation models have greatly promoted the development of the community. The transformer is the current key model and most recent participants are using a big- setting transformer model to improve the quality of translations. However, developing a more efficient model is also important. We here use a smaller model and limited computation resources to pursue high-quality translation models. Our systems are based on the Transformer model with base settings, and the models are trained on the parallel corpus of Japanese and English (Morishita et al., 2019). We compared the quality of translations by using fine-tuning with several datasets. Also, we tested several different hyperparameters of the training to find suitable values for the task. After the fine-tuning, we tried to perform ensembling of multiple results from the model to earn a better-quality translation in the English to/from Japanese model. Here we describe the details of our systems. 2 Data selection and preprocessing We select a suitable parallel corpus for model fine- tuning. We compare WMT provided dataset (which contained 7 different sources including the JParaCrawl dataset), KFTT (Kyoto Free Translation Task data set, Neubig, 2011), the JParaCrawl dataset (ver 2) and so on. We performed fine-tuning for these datasets and found that the model trained on the JParaCrawl dataset achieved better performance. We used a test data set made from WMT provided data and compare model performances by BLEU score. The score of the no fine-tuned model was 37.21, KFTT fine- tuned model was 14.87 and JParaCrawl fine-tuned model was 44.09. Therefore, we decided to use JParaCrawl as our fine-tuning dataset finally. We also consider that JParaCrawl has a reasonable amount of data for our high-efficiency training strategies. Before we use the dataset, we check the corpus data to clean up. The JParaCrawl dataset contains over 10 million sentence pairs which were constructed by broadly …

Content truncated to 500 words