TSU HITS's Submissions to the WMT 2024 General Machine Translation Shared Task

Vladimir Mynka, Nikolay Mikhaylovskiy

Abstract: This paper describes the TSU HITS team’s submission system for the WMT’24 general translation task. We focused on exploring the capabilities of discrete diffusion models for the English-to-{Russian, German, Czech, Spanish} translation tasks in the constrained track. Our submission system consists of a set of discrete diffusion models for each language pair. The main advance is using a separate length regression model to determine the length of the output sequence more precisely.