![\"Extracting](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12759647%2Fa55560042d5bd16af85ce327fd482d5a%2FScreenshot%202026-03-24%20125428.png?generation=1774337100974375&alt=media\")
I started working on this competition in the last 15 days. I found AKT PDFs data could be useful, so I used some code of this notebook's to extract tables from each page.
\nExtracting Table\n
I then created PDFs with just tables using this code.
\nProcessed PDF View\n![\"Processed](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12759647%2F32ad3aec20991b64a2d0858dc94b5244%2FScreenshot%202026-03-24%20125219.png?generation=1774336958461535&alt=media\")
I then OCRed these processed tabular PDFs using PaddleOCR, it was the best OCR method as per this paper
\nUploaded it as Raw Dataset here
\nI tried directly OCRing the pdfs but PaddleOCR had quite difficulty in recognising the tables directly.
\nI did a lot of pre-processing to make it good enough for training but did some mistakes as well due to lack of time.
\nSequence: Quick Clean -> Fix NaNs -> Fix Symbols -> Final Fixes
\nComplete solution write-up under way!
", "messages": [ { "id": "3427518", "postDate": "03/24/2026 07:22:52", "content": "I started working on this competition in the last 15 days. I found AKT PDFs data could be useful, so I used some code of this notebook's to extract tables from each page.
\nExtracting Table\n
I then created PDFs with just tables using this code.
\nProcessed PDF View\n
I then OCRed these processed tabular PDFs using PaddleOCR, it was the best OCR method as per this paper
\nUploaded it as Raw Dataset here
\nI tried directly OCRing the pdfs but PaddleOCR had quite difficulty in recognising the tables directly.
\nI did a lot of pre-processing to make it good enough for training but did some mistakes as well due to lack of time.
\nSequence: Quick Clean -> Fix NaNs -> Fix Symbols -> Final Fixes
\nComplete solution write-up under way!
", "rawMarkdown": "I started working on this competition in the last 15 days. I found AKT PDFs data could be useful, so I used some code of [this notebook's](https://www.kaggle.com/code/angantyr/text-extraction-from-pdf-docs-using-glm-oc) to extract tables from each page.\n\nExtracting Table\n\n\nI then created PDFs with just tables using [this code](https://www.kaggle.com/code/paritoshtripathi5/akkadian-pdf-layout-analysis-extract-tables). \n\nProcessed PDF View\n\n\nI then OCRed these processed tabular PDFs using PaddleOCR, it was the best OCR method as per [this paper](https://arxiv.org/pdf/2601.21957)\n\nUploaded it as Raw Dataset [here](https://www.kaggle.com/datasets/paritoshtripathi5/raw-akt-akkadian-dataset/data)\n\nI tried directly OCRing the pdfs but PaddleOCR had quite difficulty in recognising the tables directly.\n\nI did a lot of pre-processing to make it good enough for training but did some mistakes as well due to lack of time.\n\nSequence: [Quick Clean](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-preprocessing) -> [Fix NaNs](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-nan-fixes) -> [Fix Symbols](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-symbols-fix/) -> [Final Fixes](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-more-symbols-fix/)\n\nComplete solution write-up under way!", "votes": null }, { "id": "3427524", "postDate": "03/24/2026 07:37:00", "content": "I'm glad you found the tool useful!
\nThe layout reconstruction was one of my biggest issues. Now I wonder how far it could go given the modified OCR LLM model.
\nThank you for sharing your insights and notebook solutions.
", "rawMarkdown": "I'm glad you found the tool useful!\n\nThe layout reconstruction was one of my biggest issues. Now I wonder how far it could go given the modified OCR LLM model.\n\nThank you for sharing your insights and notebook solutions.", "votes": null } ], "comments": [ { "id": 3427524, "author_name": "angantyr", "author_url": "", "post_date": "03/24/2026 07:37:00", "content": "I'm glad you found the tool useful!
\nThe layout reconstruction was one of my biggest issues. Now I wonder how far it could go given the modified OCR LLM model.
\nThank you for sharing your insights and notebook solutions.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3427518": "I started working on this competition in the last 15 days. I found AKT PDFs data could be useful, so I used some code of [this notebook's](https://www.kaggle.com/code/angantyr/text-extraction-from-pdf-docs-using-glm-oc) to extract tables from each page.\n\nExtracting Table\n\n\nI then created PDFs with just tables using [this code](https://www.kaggle.com/code/paritoshtripathi5/akkadian-pdf-layout-analysis-extract-tables). \n\nProcessed PDF View\n\n\nI then OCRed these processed tabular PDFs using PaddleOCR, it was the best OCR method as per [this paper](https://arxiv.org/pdf/2601.21957)\n\nUploaded it as Raw Dataset [here](https://www.kaggle.com/datasets/paritoshtripathi5/raw-akt-akkadian-dataset/data)\n\nI tried directly OCRing the pdfs but PaddleOCR had quite difficulty in recognising the tables directly.\n\nI did a lot of pre-processing to make it good enough for training but did some mistakes as well due to lack of time.\n\nSequence: [Quick Clean](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-preprocessing) -> [Fix NaNs](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-nan-fixes) -> [Fix Symbols](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-symbols-fix/) -> [Final Fixes](https://www.kaggle.com/code/paritoshtripathi5/akkadian-akt-dataset-more-symbols-fix/)\n\nComplete solution write-up under way!", "3427524": "I'm glad you found the tool useful!\n\nThe layout reconstruction was one of my biggest issues. Now I wonder how far it could go given the modified OCR LLM model.\n\nThank you for sharing your insights and notebook solutions." }, "source": "meta" }