Robustness of Zero-Shot Cross-Lingual Performance in Multimodal vs. Text-Only Models with Adversarial Intermediate Tasks

Intermediate-task training---fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task---often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tas

Research goal: How does the robustness of zero-shot cross-lingual performance on XTREME-R vary between multimodal models (e.g., PaLI) and text-only models (e.g., XLM-R) when using adversarial intermediate tasks (e.g., BoolQ) versus standard intermediate tasks?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 8.7/10.