Keep newlines in detected texts in Azure Form Recognizer

Cyril Carraz 31

Hello,

I'm training a custom neural model in Azure Form Recognizer v3.0 and there are certain fields in the annex that are long texts (made of many paragraphs, bullet lists, etc...), these texts may span accross multiple pages.

The model can detect these texts easily, but it seems to sometimes (and only sometimes) remove the newline characters (\n) in the value/content of the field. How do I avoid that happening ?

I tried relabeling these texts in a table with a single column to make each line in a single row and join them with "\n" in the postprocessing step later, but that made the detection much worse, sometimes it even skips certain lines or paragraphs (or worse, it skips certain dates for some reason).

How do I parse the text and keep it exactly as is line by line ?

romungi-MSFT 43,451 Reputation points Microsoft Employee

2023-02-20T17:04:48.5733333+00:00

@Cyril Carraz Are you seeing the paragraphs field in your analyze response? This usually keeps the entire text as a logical unit in the response and it is easier to read from a document. The pre-built models like Read and Document are well trained to handle scenarios where the paragraphs span multiple pages. You can try these from the studio with the sample pages or upload your document and check if this model works better. However, the paragraph units are seen as a whole unit and spacing between lines as seen visually in the document is not captured.

For custom models labeling the entire text using shift select method is a good idea but there are some known limitations with page boundaries for custom neural models if paragraphs span multiple pages.
Luis Garreta 0 Reputation points

2023-06-01T14:15:06.36+00:00

Also, it happens to me with a "special" invoice type that contains several multi-line regions. However, the recognizer detects the lines in the OCR results as they appear in result.page.lines. But, I don´t know how to associate these lines with the current field with multilines.
Luis Garreta 0 Reputation points

2023-07-21T18:56:14.0166667+00:00

To get text with new lines you can create a python script that search for newlines in the OCR text and add them in the same positions in the content text of each field.

Share via

Keep newlines in detected texts in Azure Form Recognizer