SCIR-MT's Submission for WMT24 General Machine Translation Task

Baohang Li, Zekai Ye, Yichong Huang, Xiaocheng Feng, Bing Qin

Abstract: This paper introduces the submission of SCIR research center of Harbin Institute of Technology participating in the WMT24 machine translation evaluation task of constrained track for English to Czech. Our approach involved a rigorous process of cleaning and deduplicating both monolingual and bilingual data, followed by a three-stage model training recipe. During the testing phase, we used the beam serach decoding method to generate a large number of candidate translations. Furthermore, we employed COMET-MBR decoding to identify optimal translations.