The article details the collaboration between Unbabel and Instituto Superior Técnico for the WMT 2023 Shared Task on Quality Estimation (QE). Their team participated in all tasks, including sentence and word-level quality prediction and fine-grained error span detection. The team used the COMET KIWI-22 model, which showed significant improvements in correlation with human judgements. The multilingual approaches used ranked first for all tasks, surpassing the second-best multilingual submission. The team also publicly released two of their best models, COMET KIWI-XL and -XXL, which are the largest open-source QE models publicly released to date.
Publication date: 22 Sep 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2309.11925
