Simultaneous Machine Translation
Simultaneous machine translation (SiMT) aims to generate target language translations in real-time as source language input is received, balancing translation quality with minimal latency. Current research focuses on developing adaptive read/write policies, often integrated with large language models (LLMs) or employing novel architectures like non-autoregressive transformers, to optimize this trade-off. These advancements are improving the accuracy and efficiency of SiMT, with implications for real-time communication technologies such as simultaneous interpretation and speech translation. The field is also actively exploring improved evaluation metrics that better capture the nuances of real-time translation performance.
Papers
On the Hallucination in Simultaneous Machine Translation
Meizhi Zhong, Kehai Chen, Zhengshan Xue, Lemao Liu, Mingming Yang, Min Zhang
A Non-autoregressive Generation Framework for End-to-End Simultaneous Speech-to-Speech Translation
Zhengrui Ma, Qingkai Fang, Shaolei Zhang, Shoutao Guo, Yang Feng, Min Zhang
Agent-SiMT: Agent-assisted Simultaneous Machine Translation with Large Language Models
Shoutao Guo, Shaolei Zhang, Zhengrui Ma, Min Zhang, Yang Feng