Bert as an OCR metric about bert_score HOT 3 CLOSED

I am not sure I follow but hey, feel free to give it a try.

Most of the backend of BERTScore is in https://github.com/Tiiiger/bert_score/blob/master/bert_score/utils.py
You can reuse the code here to get the embeddings and pairwise cosine similarity.

I am closing this issue for now because it is less relevant to the most common usage of this repo. However, feel free to reach out if you have any question about the code.

from bert_score.

hi @gaceladri , I am not very familiar with OCR and how you evaluate it.

I am happy to chat more if you can help provide more context/background.

from bert_score.

@Tiiiger Well, it's a bit tricky, since, depending on the quality of your OCR model, the outputs can be more or less clear.

For example if you don't have a supervised corpus that it use to be the case, you can not measure the performance of the model, normally already pre-trained and you don't use to have the supervised corpus.

So my thinking is, can we measure the similarity of the outputs from the OCR and do... a cosine similarity? for example to some dictionary (the embedding of a BERT) and measure the performance of the OCR or directly replace those noisy outputs from the OCR model to the nearest words in the embedding.

Do you know what I mean?

from bert_score.