Latin NLP Tools Comparison

This project evaluates the speed, accuracy, and usability of four prominent Natural Language Processing (NLP) tools for Latin texts:

• CLTK (Classical Language Toolkit)
• Stanza for Latin (Stanford NLP)
• UDPipe
• TreeTagger

Two samples are used for testing:

• Early medieval glosses from the Vienna Bede
• Early medieval texts from the Late Latin Charter Treebank

Project Goal

To provide a reproducible and comparative analysis of Latin NLP tools for tokenisation, lemmatisation, and POS-tagging, as well as processing speed.

Project Structure

• data/: Sample Latin texts (raw and preprocessed)
• notebooks/: Jupyter notebooks for experiments
• scripts/: Python scripts for preprocessing and tool execution
• results/: Accuracy/speed metrics and visualizations

Installation

1. Clone the repo:

   git clone https://github.com/YOUR_USERNAME/latin-nlp-comparison.git
   cd latin-nlp-comparison ```

2. Create a virtual environment:

    source env/bin/activate ```

3. Install dependencies:

pip install -r requirements.txt

Metrics Evaluated

• Accuracy
  • Tokenisation
  • Lemmatisation
  • POS
• Speed: length of time to process data
• Usability: Observational assessment of set-up complexity, packages required, interface, export options

Wiki

See the GitHub Wiki for documentation, tool setup guides, and detailed findings.

Acknowledgments

Supervised by Bernhard Bauer at the University of Graz.