README.md

metadata

title: Czech GEC Punctuation Pipeline
emoji: 🔄
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: apache-2.0

🇨🇿 Czech Grammar & Punctuation Correction Pipeline

A comprehensive pipeline that combines grammatical error correction with punctuation restoration for Czech text.

🎯 Features

Pipeline Mode

• Two-Stage Processing:
  1. Grammar correction using ByT5 Czech GEC model
  2. Punctuation and capitalization using XLM-RoBERTa
• 9 Output Variants: Each stage produces 3 variants (Conservative, Balanced, Exploratory)
• Visual Grid Layout: Easy comparison of all combinations

Benchmark Mode

• Ground Truth Comparison: Test how well the pipeline recovers original text
• Similarity Scoring: Percentage match for each variant
• Comma Removal Tool: Quick test case generation
• Best Match Identification: Automatically highlights the closest match

🔧 How It Works

1. Input → Czech text with errors and missing punctuation
2. Stage 1: GEC → 3 grammatical corrections
3. Stage 2: Punctuation → 3 punctuation variants for each GEC output
4. Output → 9 total combinations to choose from

📊 Benchmark Workflow

1. Paste your original, correct Czech text
2. Click "Remove Commas" to create a test case
3. Manually introduce grammatical errors
4. Run benchmark to see recovery accuracy
5. Review similarity scores and best match

🚀 Models Used

• Grammar Correction: ufal/byt5-large-geccc-mate

  • ByT5-large model fine-tuned on Czech GEC corpus
  • Handles complex grammatical errors

• Punctuation: kredor/punctuate-all

  • Token classification model for punctuation restoration
  • Supports Czech and 11 other languages
  • Adds punctuation marks: . , ? - :

💡 Use Cases

• Text Correction: Fix both grammar and punctuation in one pipeline
• Quality Assessment: Benchmark correction models against known good text
• Model Comparison: Compare different correction strategies
• Educational Tool: Understand how different parameters affect output

🎨 Generation Strategies

Conservative

• Minimal changes, high confidence
• Lower beam search, no sampling
• Best for minor corrections

Balanced

• Moderate corrections
• Mixed parameters
• Good general-purpose option

Exploratory

• More creative corrections
• Higher diversity
• Best for heavily corrupted text

📈 Performance

• Processing time: 5-15 seconds for full pipeline
• Best on GPU but works on CPU
• Memory requirement: ~16GB for both models

🔍 Example Input

včera jsem šel do obchodu a koupil jsem si rohlíky máslo a mléko bylo to levné

This will be corrected for grammar and have punctuation added in 9 different ways!