[Discussion] What's the best way for matching OCR text?

[hv0905]

Currently we use BERT model (more precisely, bert-base-chinese) to vectorize OCR text, then use COSINE distance for indexing and searching.

However, this method seems to have low performance when processing partial keywords or semantically similar sentences.

For instance,

FYI, the OCR text of the image:
1. please
2. 你最
3. 叔
4. 什么情况兄弟
5. 爱
6. 爱
7. 害怕
8. 乳
9. 嘿

And only when I provide more detailed text, the server can return some more accurate result:

Any solution to improve the OCR text matching?

Related code

https://github.com/hv0905/NekoImageGallery/blob/master/app/Services/transformers_service.py#L59

Related documentation

https://huggingface.co/tasks/sentence-similarity