A coreference resolution pipeline uses English-to-target machine translation with cycle-consistency validation: back-translated samples are scored via BERT cosine similarity, and those scores weight training loss. Tested on four low-resource languages, the approach expands training data without manual annotation.
HOW THIS AFFECTS YOU
●
researcherThe cycle-consistency weighting via BERT cosine similarity is a reusable technique for filtering noisy MT-generated labels in other low-resource NLP tasks.