Extracting text from book PDFs using machine learning

What we’ve done so farbook scanned PDF

• Textualized versions of books contain page numbers, footnotes, and various other strings.
  • Initially, I thought line continuation judgment would identify whether text is directly connected or connected with intervening whitespace, but I was wrong.
• Need a mechanism to identify “what is this” on a line by line basis for information on one page - Page Row-Based Language Models
  • At first, I was thinking of creating a language model-like form of learning, but when I wrote a rule-based model as a sample, it was surprisingly easy to create.
• This worked well when I tried it in “The Engineer’s Art of Intellectual Production”, but when I used it in “The Art of Exploration of Knowledge”, I didn’t get any text at all.
  • The former is indented with full-width spaces at the beginning of paragraphs, while the latter is not, so
• Rather than modify the rule-based version, we decided to make it machine learning.
  • I made one line as a baseline, assuming that the input would be five lines, plus two lines before and after, and it worked well enough. - OCR Garbage Cleaning may be required separately
• Added on 2018-10-18:.
  • The idea of “5 lines with a line before and after” is like textCNN, but Transformer is more interesting in many ways.
  • The row feature vector is not made from the word embedded vector, similarly the page feature vector should be made - Automatic generation of table of contents
• Convert text converted from PDF to feature vector as one item per line (output-features)
  • Set of 4 lines: rows, features, labels, and blank lines
• A human looks at it and assigns a label as appropriate.
• Active-learning based on the learning data
  • Output “labeled features”, “their labels”, and “features that are not confident in their identification” in json
    • It’s active learning for multi-class classification, so it’s sorted by the difference between first and second place.
  • learned model
• Merge
  • Merge the original data with the training result data.
  • In fact, you can add new features here, since they are newly re-created without using the output-features feature.
  • Use unique pageRatio, lineRatio pairs to link to existing data
  • Output merged results to honbun_merged_features.txt, diff to make sure there are no major problems, then manually rename to honbun_features.txt
• Main text output (output-honbun)
  • Extract the main text from the learned model and the feature data.

Minimum Procedure

• —output-feature to create feature vector
• —output-honbun —use-model to specify intellitech/honbun_model.npz
• Look at the resulting honbun.txt file, and if it looks OK, OK.

---

This page is auto-translated from /nishio/機械学習による書籍PDFからの本文抽出 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.