Improving Simultaneous Machine Translation with Monolingual Data

Simultaneous machine translation (SiMT) is usually done via sequence-level knowledge distillation (Seq-KD) from a full-sentence neural (NMT) model. However, there still significant performance gap between NMT and SiMT. In this work, we propose to leverage monolingual data improve SiMT, which trains SiMT student on the combination of bilingual external distilled by Seq-KD. Preliminary experiments En-Zh En-Ja news domain corpora demonstrate that can significantly quality (e.g., +3.15 BLEU En-Zh). Inspired behavior human simultaneous interpreters, novel sampling strategy for considering both chunk length monotonicity. Experimental results show our consistently outperforms random (and other conventional typical strategies) avoiding key problem -- hallucination, has better scalability. We achieve +0.72 improvements average against En-Ja. Data codes be found at https://github.com/hexuandeng/Mono4SiMT.