SimplifyAI - Offline Text Accessibility Tool

Privacy-preserving, AI-powered text simplification for neurodivergent readers

A desktop application that uses fine-tuned T5 transformer models to simplify text for dyslexia, ADHD, and autism accessibility needs. Runs completely offline on your local CPU—no cloud, no data sharing, no internet required.

---

🚀 Quick Start

Desktop App (Recommended)

# Install dependencies
npm install

# Start the application
npm run start

This launches the Electron app with Vite dev server. For production build, use npm run build.

Command Line

# Install Python dependencies
pip install transformers torch spacy
python -m spacy download en_core_web_sm

# Run simplification
python simplify.py --input sample.txt --mode dyslexia --metrics

---

✨ Features

Three Accessibility Modes

Mode	Best For	Key Features
Dyslexia	Readers with dyslexia	Short sentences, visual spacing, optional hyphenation, text-to-speech
ADHD Focus	Attention maintenance	Progress markers [1/N], bolded key nouns, focus navigation
Literal Clarity	Autism spectrum	Idiom replacement, jargon simplification, literal interpretations

Privacy & Performance

• 🛡️ 100% Offline – All processing happens locally on your CPU
• ⚡ Fast – T5-small model runs efficiently on modern CPUs (~1-2 sec per paragraph)
• 🔒 No Data Leaves Your Device – Perfect for sensitive documents
• 📦 Self-contained – No external API keys or cloud dependencies

Accessibility Settings

• Themes: Light, Dark, High Contrast (WCAG AAA)
• Typography: Choose from Lexend, OpenDyslexic, Merriweather, or Monospace
• Layout: Adjustable font size (14-28px), line spacing (1.5-2.0)
• Motion: Reduce motion toggle for animations
• Hyphenation: Optional for dyslexia (per BDA guidelines)

---

🏗️ How It Works

Input Text → Preprocessing → Neural Simplification → Mode Formatting → Output
     ↓              (Cleanup)           (T5 + ACCESS)      (Rules-based)      ↓
  Raw text   (Idioms/Jargon/Homophone) Sentence-by-sentence   Dyslexia/ADHD/Autism  Ready-to-read

1. Neural Simplification (T5 Transformer with ACCESS Control Tokens)

   • Fine-tuned model (t5/) on accessibility-focused parallel corpora
   • Preprocessing: Idiom/jargon replacement and homophone correction
   • Sentence-level processing with ACCESS control tokens for mode-specific guidance
   • Dynamic length constraints prevent over-summarization
   • Handles spelling, grammar, and vocabulary simplification

2. Rule-Based Post-Processing

   • Dyslexia: Split on conjunctions, ensure punctuation, optional hyphenation
   • ADHD: Bold first noun per sentence, add [1/N] markers
   • Autism: Replace idioms with literal meanings, simplify jargon

3. Readability Metrics

   • Word count, sentence length, Flesch Reading Ease
   • Before/after comparison for measurable improvements

---

🎯 Mode Details

Dyslexia Mode

• Breaks complex sentences into short, single-idea statements
• One sentence per line with extra spacing
• Optional hyphenation for long words (disabled by default per BDA guidelines)
• Built-in text-to-speech with word boundary tracking
• Highlights paragraph as it reads aloud

ADHD Focus Mode

• Converts paragraphs into a numbered bullet list: [1/15] - Sentence...
• Visual emphasis: First significant noun in each sentence is bolded
• Uses spaCy POS tagging for accurate noun detection (with fallback heuristics)
• Keyboard navigation: arrow keys to cycle through sentences
• Active sentence highlighting in the center of the viewport

Autism/Literal Clarity Mode

• Idiom replacement: "break a leg" → "good luck"
• Jargon simplification: "PTSD" → "Post-Traumatic Stress Disorder"
• Context awareness: Prevents over-replacement
  • Example: "He broke his leg" (literal injury) is NOT replaced with "good luck"
• 5,200+ idioms and 2,000+ jargon terms (morphologically expanded)
• Bolded replacements for easy identification

---

📦 What's Included

offline-text-accessibility/
├── simplify.py           # Core simplification engine (CLI)
├── simplify_server.py    # Persistent server for Electron
├── dyslexia_mode.py      # Dyslexia formatter
├── adhd_mode.py          # ADHD focus formatter
├── autism_mode.py        # Literal clarity formatter
├── utils.py              # Utilities: sentence splitting, metrics, homophones
├── nlp_utils.py          # spaCy integration (POS tagging, noun extraction)
├── idiom_map.json        # 5,207 idiom → literal mappings
├── jargon_map.json       # 2,054 jargon → plain language mappings
├── electron-app/         # Desktop UI (React + Electron)
│   ├── main.js
│   ├── preload.js
│   └── src/
│       ├── App.jsx
│       ├── context/
│       └── components/
├── t5/        # Fine-tuned T5 model (local)
├── expand_maps.py        # Utility: expand maps with morphological variants
├── __tests__/            # Unit tests
└── docs/                 # Additional documentation

---

🔧 Installation

Desktop App (Electron)

# Clone the repository
git clone https://github.com/yourusername/offline-text-accessibility.git
cd offline-text-accessibility

# Install Node.js dependencies
npm install

# (Optional) Install Python dependencies for backend
pip install transformers torch spacy
python -m spacy download en_core_web_sm

# Run the app
npm run start

Python CLI Only

pip install transformers torch spacy
python -m spacy download en_core_web_sm

---

🖥️ Using the Desktop App

1. Paste or type your text in the left panel
2. Select a mode: Dyslexia, ADHD Focus, or Literal Clarity
3. Click "Simplify Text"
4. View results in the right panel, customized for the selected mode
5. Adjust settings (gear icon):
   • Change theme (light/dark/high-contrast)
   • Select font (Lexend, OpenDyslexic, etc.)
   • Adjust font size and line spacing
   • Enable/disable hyphenation (dyslexia)
   • Reduce motion (accessibility)

---

📊 Readability Metrics

The tool calculates and displays:

• Word count – before and after simplification
• Average sentence length – target: shorter is better
• Flesch Reading Ease – 0-100 score (higher = easier to read)

All changes are color-coded: green for improvement, red for regression.

---

🛠️ For Developers

Architecture

The application consists of three layers:

1. Frontend – React UI with Electron shell

   • State managed via React Context (AppContext, ThemeContext)
   • IPC communication with Python backend
   • TailwindCSS for styling with CSS custom properties

2. Backend – Python simplification server

   • T5 transformer model (HuggingFace) with ACCESS control token guidance
   • Preprocessing pipeline: Idiom/jargon replacement and homophone correction
   • Sentence-level processing with dynamic length constraints to prevent over-summarization
   • LRU model cache (up to 3 models)
   • JSON-over-stdin/stdout protocol

3. Models – Fine-tuned T5 on accessibility data

   • Location: t5/ (local directory)
   • ACCESS control token profiles for mode-specific guidance
   • Dynamic length constraints and penalties for optimal output
   • Tuned generation parameters (beam search, repetition penalty)
   • Preprocessing-ready design for clean input

Model Selection

# In simplify.py
model_choice = "small"      # ./t5 (default, fastest)
# or
model_choice = "medium"     # t5-medium (higher quality, more RAM)
# or
model_choice = "auto-task"  # Auto-select based on text complexity
# or
model_choice = "auto-device" # Auto-select based on available RAM

---

📝 Recent Updates

• Expanded lexical resources: Idiom map grown from 220→5,207 entries; jargon map from 100→2,054 entries via morphological expansion
• Full mode implementation: All three accessibility modes now fully functional
• spaCy integration: Accurate part-of-speech tagging for ADHD noun bolding
• Improved sentence splitting: Better handling of abbreviations, decimals, and titles
• Context-aware homophone correction: "hole" vs "whole", "knot" vs "not", etc.
• Comprehensive test suite: Unit tests for mode formatters and map integrity
• Detailed documentation: Architecture deep dive and developer guide

---

⚠️ Limitations

• English text only (models trained on English)
• Requires ~2GB RAM for T5-medium (smaller model uses less)
• Not a clinical tool – designed for general accessibility support
• No user studies yet – feedback welcome

---

📄 License

MIT

---

🙏 Acknowledgments

• T5 model and HuggingFace transformers library
• spaCy for natural language processing
• TailwindCSS for utility-first styling
• The neurodiversity community for feedback and inspiration

---

📚 Documentation

• Architecture Deep Dive – Comprehensive technical overview
• CLAUDE.md – Development guide and project instructions
• docs/ – Additional evaluation and planning documents

---

Made with ❤️ for neurodivergent readers. All processing happens on your device. Your text, your privacy.