Code from https://github.com/FreddeFrallan/Multilingual-CLIP. Modified the embedding dimension to match the dimension for mBERT for comparison.
- Enlgish-Chines/Chinese-English
- German-English/English-German
- Russian-English/English-Russian
WMT 19 dataset from huggingface
- Very initial experiments:
- The pre-training for mBERT and mCLIP is not using the same dataset
- cannot really do the vector multiplication because the vector length is not aligned