bug/Incorrectly classifying Chinese text as "title"

Hello unstructured team,

I am currently using unstructured with paddle ocr to try and pre-process chinese documents.

The goal is to strip away any paratext in the document and only keep main body text.

The problem is all the text is either falsely classified as "title" or as "uncategorizedtext."

I am using:

paddeocr>-2.9.1 (not sure if this is being utilized at all by unstructured)
paddlepaddle==3.0.0b1
pymupdf>=1.25.3
unstructured-ingest>=0.5.5
unstructured-paddleocr==2.8.1.0
unstructured[pdf]>=0.16.21

I saw in the release logs that this problem was addressed in unstructured v 0.16.12.

The code I am using is the example code here: https://docs.unstructured.io/open-source/how-to/set-ocr-agent

But instead of importing partition_image from unstructured.partition.image, I am importing partition_pdf from unstructured.partition.pdf.

Please let me know if I am using your library correctly, let me know if there are any hyperparameters that I can set to improve the classification, let me know if this is just currently a limitation - does the API do a better job?

Thanks in advance.

Start with the OCR-agent example linked in the issue and the partition_pdf entry point, then compare the reported PaddleOCR, PaddlePaddle, PyMuPDF, unstructured-ingest, and unstructured versions with the 0.16.12 release notes. Determine whether the Chinese classification is an expected limitation, a configuration issue, or a reproducible bug; done means the cause and supported next step are established.