Humanize Research Paper Text (texthumanize CLI)

This mini-project lets you take a full research-paper-style .txt file (for example, exported from a paper editing site) and run it through the texthumanize library using its academic profile.

The script:

• Preserves the overall formatting of your text file (headings, spacing, etc.) while normalizing the language.
• Uses humanize_chunked() so very long papers can be processed safely.
• Is configured for formal academic style by default (profile="academic").

Important limitation: Neither this script nor texthumanize can guarantee that your text will evade all external AI detectors (GPTZero, Turnitin, Originality.ai, etc.). The library is designed to reduce common AI-style markers and normalize style, but external detectors use their own proprietary models and signals.

---

1. Install dependencies

From the py-humanize folder:

pip install -r requirements.txt

This installs:

• texthumanize ≥ 0.27.1
• lxml (used for safe DOCX XML edits)

---

2. Prepare your input file

Place your formatted research paper text in a plain UTF‑8 .txt file, for example:

• whole.txt

You can keep the editing site's template structure in this file; the script will humanize the content as raw text while preserving line breaks and most structural formatting.

---

3. Run the humanizer

Basic usage (from the py-humanize directory):

python humanize_paper.py whole.txt

This will:

• Read whole.txt
• Humanize it with profile="academic" and English language (lang="en")
• Write the result to whole.txt.humanized.txt in the same directory

---

4. Advanced options

You can customize behavior via CLI flags:

python humanize_paper.py whole.txt \
  -o whole_academic_humanized.txt \
  --lang en \
  --profile academic \
  --intensity 30 \
  --seed 42 \
  --chunk-size 3000

• -o / --output: Custom output path.
• --lang: Language code (default: en).
• --profile: Processing profile (default: academic). Other options in texthumanize include chat, web, seo, docs, formal, marketing, social, email.
• --intensity: 0–100. If omitted, texthumanize uses the profile-specific default (for academic, ~25).
• --seed: Ensures deterministic output for the same input.
• --chunk-size: Approximate characters per chunk when splitting long documents internally.

---

5. Notes on AI detection

• texthumanize focuses on style normalization (sentence length, connectors, bureaucratic phrases, etc.).
• It can reduce AI-style markers, but the project itself explicitly states it cannot guarantee bypassing all third‑party AI detectors.
• For best results:
  • Start with profile="academic" and low–moderate intensity (20–40).
  • Manually review the humanized text to ensure scholarly tone, correct terminology, and no template artefacts were altered in an unintended way.

---

DOCX Support (Word .docx)

This project also includes a DOCX-focused script: humanize_docx.py.

What it edits

• Only the main body text inside word/document.xml by replacing the contents of <w:t> text nodes.
• It skips:
  • Word field instructions (w:instrText)
  • Word math nodes (m:*)

Pitfalls / complications

• Word often splits visible text across multiple runs/text nodes, so humanizing per <w:t> can reduce global coherence.
• Formatting is preserved structurally (we do not change w:rPr, paragraph/run structure, etc.), but the wording changes are still inserted at the text-node level.

Usage

python humanize_docx.py input.docx -o output.docx

Optional flags:

python humanize_docx.py input.docx -o output.docx \
  --lang en \
  --profile academic \
  --intensity 30 \
  --seed 42 \
  --min-chars 2

---

Repairing DOCX Outputs (newline cleanup)

If you see odd reflow/layout issues after humanize_docx.py, it can help to remove newline characters that end up inside <w:t> text nodes.

Run:

python repair_docx_wt_newlines.py input.docx -o output.docx

This edits only word/document.xml and validates that no <w:t> nodes still contain \n or \r after repair.

---

PDF Support (.pdf)

This project also includes a PDF-focused script: humanize_pdf.py.

What it does

• Uses PyMuPDF to extract embedded text along with bounding boxes.
• Humanizes the extracted text line-by-line using texthumanize.
• Covers the original text regions with a white rectangle and re-inserts the humanized text back into the same approximate regions.

Important limitations

• This works best for PDFs that contain real, extractable text (not scanned images).
• Even for text-based PDFs, fonts/colors/spacing may shift slightly depending on how the PDF stores glyph positioning.

Usage

python humanize_pdf.py input.pdf -o output.pdf

Optional flags:

python humanize_pdf.py input.pdf -o output.pdf \
  --lang en \
  --profile academic \
  --intensity 30 \
  --seed 42 \
  --min-chars 2 \
  --min-text-chars 50