texthumanizer 1.1.0

Offline AI-text humanizer that preserves images, tables, and research context.

texthumanizer 📝

Offline AI-text humanizer & plagiarism reducer for students and researchers. No internet needed after model download. Preserves research context, citations, and semantic meaning.

---

✨ Features

Feature	Detail
Humanize AI text	Rewrites ChatGPT / Claude / Gemini output to sound natural
Plagiarism reduction	Paraphrase-based, not just synonym swap
Semantic preservation	Meaning, tone, and argument structure kept intact
Research-aware	Citations [1], abbreviations DNA, units 95%, et al. — all preserved
DOCX support	Paragraph-level processing; headings untouched
100% offline	T5-based model runs locally after first download (~250 MB)
Lightweight	CPU-friendly, no GPU required

---

📦 Installation

Full install (recommended)

pip install texthumanizer[all]

Minimal (text only, no docx)

pip install texthumanizer[ml]

With DOCX support only

pip install texthumanizer[ml,docx]

First run downloads the T5 model (~250 MB) from HuggingFace once and caches it locally.

---

🚀 Quick Start

1. Humanize pasted text

from texthumanizer import TextHumanizer

th = TextHumanizer()

ai_text = """
Artificial intelligence has rapidly transformed numerous sectors of society,
demonstrating unprecedented capabilities in natural language processing,
computer vision, and decision-making systems.
"""

result = th.humanize_text(ai_text)
print(result)

2a. Humanize a .docx → save new .docx

from texthumanizer import TextHumanizer

th = TextHumanizer()

# Saves "humanized_my_essay.docx" next to the original
output_path = th.humanize_doc("my_essay.docx", output="doc")
print(f"Saved: {output_path}")

# Custom output path
th.humanize_doc("my_essay.docx", output="doc", output_path="D:/final_essay.docx")

2b. Humanize a .docx → get plain text back

from texthumanizer import TextHumanizer

th = TextHumanizer()
text = th.humanize_doc("my_essay.docx", output="text")
print(text)

---

⚙️ Configuration

th = TextHumanizer(
    diversity=0.7,    # 0.0 = minimal changes, 1.0 = maximum rewriting (default: 0.7)
    device=-1,        # -1 = CPU, 0 = GPU (default: -1)
    verbose=True,     # Show progress (default: True)
)

Parameter	Range	Effect
diversity=0.3	Low	Light rewording, very safe for technical papers
diversity=0.7	Medium	Balanced — good for essays and reports ✅
diversity=0.9	High	Heavy rewriting — good for blog posts or general text

---

🖥️ CLI Usage

Interactive mode

python -m texthumanizer.cli

Direct text

python -m texthumanizer.cli text "Your AI-generated text here" --diversity 0.7

Pipe from file

cat essay.txt | python -m texthumanizer.cli text

DOCX → humanized DOCX

python -m texthumanizer.cli doc essay.docx --output doc

DOCX → print text

python -m texthumanizer.cli doc essay.docx --output text

---

🔬 How It Works

Input Text
    │
    ▼
[Mask technical terms]         ← citations, abbreviations, units, years
    │
    ▼
[Split into sentences]         ← smart splitter (handles abbreviations)
    │
    ▼
[T5 Paraphrasing model]        ← humarin/chatgpt_paraphraser_on_T5_base
    │                            temperature + top-k + top-p sampling
    ▼
[Restore masked terms]         ← [1], DNA, 2023 put back exactly
    │
    ▼
Output Text

Why T5 and not GPT-style? T5 is an encoder–decoder model trained specifically on paraphrase tasks. It is:

• Much smaller (~250 MB vs multi-GB GPT models)
• CPU-friendly and fast
• Better at preserving meaning than decoder-only models

---

📋 What Gets Preserved

Type	Example	Preserved?
Academic citations	[1], [1,2,3]	✅
Author citations	(Smith et al., 2021)	✅
Abbreviations	DNA, AI, COVID, LSTM	✅
Years	2023, 1990	✅
Percentages	95%, 3.5%	✅
Scientific units	kg, MHz, nm, kcal	✅
Figure/Table refs	Fig. 3, Table 1	✅
DOIs / URLs	doi:10.xxx, https://...	✅
Latin abbreviations	et al., e.g., i.e.	✅
Headings (in .docx)	Section titles	✅ untouched

---

🧪 Example Output

Input (AI-generated):

The utilization of machine learning algorithms has demonstrated significant efficacy in the domain of medical diagnosis, achieving accuracy rates exceeding 95% in several clinical trials [1,2].

Output (humanized):

Using machine learning methods has shown strong results in medical diagnosis, reaching accuracy levels above 95% in a number of clinical studies [1,2].

---

💡 Tips for Best Results

• Research papers: Use diversity=0.4 – 0.6 to keep technical accuracy
• Essays / assignments: Use diversity=0.7 (default)
• Blog posts / creative writing: Use diversity=0.8 – 0.9
• Process section-by-section for long papers for best control
• GPU users: set device=0 for ~5× speed improvement

---

🔬 How It Works (In-place Replacement)

Unlike other humanizers that strip away formatting, texthumanizer uses an "In-place run-level replacement" strategy:

1. It creates a temporary copy of your .docx.
2. It identifies text-bearing "runs" within each paragraph.
3. It humanizes the text while skipping runs that contain images or drawings.
4. It injects the new text back into the original XML structure, keeping your layout 100% intact.

📄 License

MIT License — free for personal and academic use.

---

⚠️ Disclaimer & Ethics

This tool is designed to assist researchers in improving the readability of their own writing. It is NOT intended for academic dishonesty or bypassing plagiarism checks for unoriginal work. Use responsibly and always cite your AI assistance if required by your institution.