How to calculate the image_text_similarity scores for both Chinese and English?

[weiaicunzai]

Thank you for your excellent work.

Regarding my dataset, which includes both English and Chinese samples, I am wondering how I can simultaneously calculate the similarity scores between image and text pairs for both languages.

[HYLcool]

Hi @weiaicunzai , thanks for your attention on Data-Juicer~

We use CLIP as the default model to calculate the embeddings of image-text pairs, which works fine on English corpus but not on Chinese texts (ref https://github.com/openai/CLIP/issues/7). For Chinese texts, models like Chinese-CLIP might perform better.

So there is a possible way to do so is to split the datasets into two subsets in English and Chinese with our dedicated dataset_split_by_language tool, and then deploy different models for the image_text_similarity_filter OP to handle them respectively.