Multilingual Coreference Resolution via Cycle-Consistent Machine Translation

Researchers propose a novel coreference resolution pipeline that uses machine translation and cycle-consistency validation to improve NLP performance in low-resource languages. By translating English training data to target languages and back-translating to verify quality, the approach generates weighted training samples that significantly enhance coreference resolution accuracy, even enabling resolution in languages without existing corpora.

This research addresses a critical gap in natural language processing where English-language models dominate while low-resource languages remain underserved. Coreference resolution—identifying when different words refer to the same entity—is fundamental to downstream NLP tasks like machine translation, question answering, and document summarization. The proposed solution leverages cycle-consistent machine translation, a technique borrowed from computer vision, to bootstrap training data quality without requiring expensive manual annotation.

The technical approach is elegant: researchers translate English coreference-annotated data into target languages, then back-translate to English and measure similarity against originals using BERT embeddings. This cosine similarity score becomes a confidence weight in the training loss function, effectively filtering out poor translations while preserving high-quality synthetic data. This reduces dependency on human-annotated corpora, which are scarce for low-resource languages.

For the AI development community, this work demonstrates how machine translation combined with cycle-consistency can democratize access to advanced NLP capabilities across linguistic boundaries. The ability to enable coreference resolution in previously unsupported languages expands the practical applicability of conversational AI, information extraction systems, and document processing tools to a global audience.

The research validates effectiveness across four low-resource languages, indicating the approach generalizes well. Future developments might combine this with multilingual models like mBERT or XLM-R to further improve performance. The methodology could potentially transfer to other NLP tasks facing similar data scarcity challenges, making it a valuable contribution to inclusive AI development.

• →Cycle-consistent machine translation enables high-quality synthetic training data generation for coreference resolution in low-resource languages.
• →Back-translation similarity scoring weights samples by quality, reducing the impact of poor MT translations on model performance.
• →The pipeline enables accurate coreference resolution in languages where no annotated corpora previously existed.
• →This approach addresses a significant gap in NLP where English-centric models dominate while other languages lag in capability.
• →The methodology generalizes across multiple languages and could transfer to other NLP tasks facing data scarcity.