Simultaneous Machine Translation using LLM on LRL (low-resource language)

Simultaneous machine translation (SiMT) allows the real-time communication, regardless of language, but its performance remains poor in low-resource languages (LRLs) because of the small amount of parallel data, morphology issues, and poor model coverage. The most recent large language models (LLMs) exhibit impressive multilingual generalization, but have not been investigated in high-latency SiMT in LRL. This paper presents a SiMT architecture based on LLM that combines incremental decoding, adaptive prefix-to-prefix signaling and optimal data augmentation to improve the quality of LRL translation, without the need to compromise low inference latency. This method will harness the synthetic parallel data, back-translation, and transfer learning between typologically related high-resource languages to overcome the data limitation. A wait-k policy is introduced as an adaptive policy to dynamically strike the balance between the translation lateness and the precision of translation as model confidence signals. Four LRLs such as Assamese, Amharic, Lao and Bhojpuri have been experimented and results indicate that the proposed structure outsmarts the baseline SiMT architecture by 19% with much less lag. The contribution of all the modules is validated through ablation analysis, whereas the case study indicates the strength of the model to code-mixing and word-order variation. The results refer to the fact that LLM-based SiMT is a potentially valuable future of scalable, real-time translation in low-resource and linguistic minority settings.