Unsupervised NMT

Unsupervised neural machine translation (UNMT) aims to train machine translation models without relying on parallel corpora, focusing instead on leveraging monolingual data and pre-trained multilingual models. Current research emphasizes addressing challenges like the "copying problem" (where the model simply replicates parts of the input) through techniques such as incorporating language discriminator losses and refined training schedules, and improving performance on low-resource languages via intermediate task fine-tuning and back-translation. These advancements, along with explorations of novel architectures like flow-adapters and the integration of emergent communication frameworks, are pushing the boundaries of UNMT, potentially enabling translation between language pairs with limited or no parallel data available.