Multilingual Intermediate Tasks for Efficient mT5 Zero-Shot Cross-Lingual Transfer

This paper studies zero-shot cross-lingual transfer of vision-language models. Specifically, we focus on multilingual text-to-video search and propose a Transformer-based model that learns contextualized multilingual multimodal embeddings. Under a zero-shot setting, we empirically demonstrate that performance degrades significantly when we query the multilingual text-video model with non-English sentences. To address this problem, we introduce a multilingual multimodal pre-training strategy, and collect a new multilingual instructional video dataset (MultiHowTo100M) for pre-training. Experimen

Research goal: Does incorporating multilingual intermediate tasks during mT5 pre-training improve its zero-shot cross-lingual transfer efficiency (inference speed and memory usage) compared to English-only pre-training when evaluated on XTREME-Z, and how does this trade-off scale with model size?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 9.0/10.