Source-Side Context Predictive Completion for Simultaneous Machine Translation

Simultaneous machine translation (SiMT) initiates the translation process while continuously receiving input from a streaming source. Unlike full-sentence machine translation, SiMT faces the inherent issue of generating target words based on only partial source input, which typically limits the quality of translation. To address this issue, we propose the novel Predictive Source Completion Framework (PSCF) to predict and complete missing source-side information, thereby enhancing SiMT performance under the same latency. In particular, PSCF improves prediction accuracy through real supervisory signals and better completes the missing information through the representation information of the predicted words. Experimental results show that PSCF improves the translation performance over the strong baselines in regard to IWSLT14 De $\rightarrow$ En, IWSLT15 En $\rightarrow$ Vi, and WMT15 De $\rightarrow$ En.