Adversarial Contrastive Training for Robust Zero-Shot Cross-Lingual Transfer in XLM-R

Existing zero-shot cross-lingual transfer methods rely on parallel corpora or bilingual dictionaries, which are expensive and impractical for low-resource languages. To disengage from these dependencies, researchers have explored training multilingual models on English-only resources and transferring them to low-resource languages. However, its effect is limited by the gap between embedding clusters of different languages. To address this issue, we propose Embedding-Push, Attention-Pull, and Robust targets to transfer English embeddings to virtual multilingual embeddings without semantic loss,

Research goal: How does the use of adversarial training with contrastive loss improve the robustness of XLM-R's zero-shot cross-lingual transfer performance, as measured by F1 scores on UDA and PAWS-X benchmarks?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 7.9/10.