AIST AIRC Systems for the WMT 2024 Shared Tasks

AbstractAt WMT 2024 AIST AIRC participated in the General Machine Translation shared task and the Biomedical Translation task. We trained constrained track models for translation between English, German, and Japanese. Before training the final models, we first filtered the parallel data, then performed iterative back-translation as well as parallel data distillation. We experimented with training baseline Transformer models, Mega models, and fine-tuning open-source T5 and Gemma model checkpoints using the filtered parallel data. Our primary submissions contain translations from ensembles of two Mega model checkpoints and our contrastive submissions are generated by our fine-tuned T5 model checkpoints.