IOL Research’s Submission for WMT 2023 Quality Estimation Shared Task

AbstractThis paper presents the submissions of IOL Research in WMT 2023 quality estimation shared task. We participate in task 1 Quality Estimation on both sentence and word levels, which predicts sentence quality score and word quality tags. Our system is a cross-lingual and multitask model for both sentence and word levels. We utilize several multilingual Pretrained Language Models (PLMs) as backbones and build task modules on them to achieve better predictions. A regression module on PLM is used to predict sentence level score and word tagging layer is used to classify the tag of each word in the translation based on the encoded representations from PLM. Each PLM is pretrained on quality estimation and metrics data from the previous WMT tasks before finetuning on training data this year. Furthermore, we integrate predictions from different models for better performance while the weights of each model are automatically searched and optimized by performance on Dev set. Our method achieves competitive results.