Orthographic Normalization for German Learner Text: Corpus Processing, Evaluation Infrastructure, and LLM Assessment
This repository contains the complete pipeline for extracting, processing, and evaluating text normalization systems on German learner corpora. It includes custom extraction tools for XML-encoded learner texts, stratified data splitting, LLM inference via API, and comprehensive evaluation scripts measuring accuracy, error reduction rate (ERR), word error rate (WER), character error rate (CER), and chrF.
Context: This work was conducted at EURAC Research (November 2025 - January 2026) to explore automation in learner text normalization evaluation, focusing on minimal-intervention orthographic correction while preserving grammatical errors.
LearnTextNorm-De/
├── corpora/ # XML source files (obtain from publishers)
│ ├── LEONIDE/
│ ├── Kolipsi_1_L1/
│ ├── Kolipsi_1_L2/
│ └── Kolipsi_2/
│
├── scripts/ # All modules and CLI tools
│ ├── extraction/ # Extraction module
│ │ ├── __init__.py
│ │ ├── config.py # Corpus configurations
│ │ ├── parsers.py # XML parsing for LEONIDE/Kolipsi
│ │ ├── sentencizer_de.py # German sentence boundary detection
│ │ ├── pipeline.py # Main extraction pipeline
│ │ └── README.md
│ │
│ ├── tsv_tools/ # TSV maintenance module
│ │ ├── __init__.py
│ │ ├── updater.py # Sync TSV with NORM edits
│ │ ├── validator.py # TSV format validation
│ │ └── README.md
│ │
│ ├── splits/ # Data splitting module
│ │ ├── __init__.py
│ │ ├── splitter.py # Stratified sampling
│ │ ├── format_converters.py # NORM/SRC/TGT/JSON generation
│ │ ├── validators.py # NORM file validation
│ │ └── README.md
│ │
│ ├── inference/ # LLM inference module
│ │ ├── __init__.py
│ │ ├── prompts.py # Zero-shot and few-shot prompts
│ │ ├── api_client.py # API wrapper (LLaMA, GPT)
│ │ ├── batch_processor.py # Batch inference with rate limiting
│ │ └── README.md
│ │
│ ├── eval/ # Evaluation module
│ │ ├── baseline.py # LAI and MFR baselines
│ │ ├── llmAlign.py # Align LLM output to original tokenization
│ │ ├── normEval.py # Accuracy and ERR computation
│ │ ├── wer++.py # WER/CER with error breakdown
│ │ └── README.md
│ │
│ ├── stats/ # Statistics module
│ │ ├── __init__.py
│ │ ├── raw_stats.py # Pre-extraction XML statistics
│ │ ├── processed_stats.py # Post-extraction TSV statistics
│ │ ├── display.py # Formatted output
│ │ └── README.md
│ │
│ ├── extraction.py # CLI: Run extraction pipeline
│ ├── tsv_update.py # CLI: Update TSV from NORM edits
│ ├── data_maker.py # CLI: Create splits and formats
│ ├── api_inference.py # CLI: Run LLM inference
│ ├── stats.py # CLI: Compute corpus statistics
│ └── README.md # Module documentation
│
├── master_files/ # Central storage
│ ├── all_corpora.tsv # Master TSV with metadata
│ └── *.norm # Individual corpus NORM files
│
├── data/ # Generated splits
│ ├── train/
│ │ ├── train.norm # Vertical format: source\target
│ │ ├── train.src # Source sentences
│ │ └── train.tgt # Target sentences
│ │ └── train_indices.tsv # Sentence Metadata (original corpus, source\target, text types, corrections)
│ ├── dev/
│ └── test/
│
├── hypos/ # Model outputs
│ ├── llama3-2/
│ │ ├── 0shot.hyp # Raw LLM output in 0 shot
│ │ └── 0shot.norm # Aligned to original tokenization
│ │ ├── 2shot.hyp # Raw LLM output in 2 shot
│ │ └── 2shot.norm # Aligned to original tokenization
│ └── gpt-oss/
│
├── .gitignore
└── README.md
Requirements:
- Python 3.8+
- spaCy with German model
- API access (optional, for LLM inference)
Note: All modules are in the scripts/ directory. Import paths use scripts.module_name format.
Dependencies:
Core requirements:
- spacy>=3.8.11 (with de_core_news_sm model)
- pandas>=2.3.5
- numpy>=1.21.0
- lxml>=4.9.0
- editdistance>=0.6.0
- tqdm>=4.67.1
Optional (for API inference):
- requests>=2.32.5
- ollama (if using ollama API)
If you use this repository or dataset, please cite:
@misc{galiero2026learntextnorm,
author = {Galiero, Lucia},
title = {Exploring Automation to Evaluate Orthographic Normalization of Learner Text: The Case of German as a L2},
year = {2026},
publisher = {EURAC Research},
howpublished = {\url{https://github.com/lgaliero/LearnTextNorm-De}}
}Data Sources:
@inproceedings{glaznieks2022leonide,
title={LEONIDE–A Longitudinal Trilingual Corpus of Young Learners of Italian, German and English},
author={Glaznieks, Aivars and Frey, Jennifer-Carmen and Stopfner, Maria and Zanasi, Lorenzo and Nicolas, Lionel},
booktitle={Proceedings of the Thirteenth Language Resources and Evaluation Conference},
pages={2486--2497},
year={2022}
}
@article{glaznieks2023kolipsi,
title={The Kolipsi Corpus Family: Resources for Learner Corpus Research in Italian and German},
author={Glaznieks, Aivars and Frey, Jennifer-Carmen and Abel, Andrea and Nicolas, Lionel and Vettori, Chiara},
journal={Italian Journal of Computational Linguistics},
volume={9},
number={1},
pages={67--86},
year={2023}
}Code and data in this repository are intended for Academic Use Only. Redistribution is not allowed
Note on Data: The original XML corpus files are not included in this repository. To reproduce the full pipeline, please download the data from their CLARIN repository page and comply by the intended use license:
- Kolipsi 1 L1 & L2 (V 1.1.): https://clarin.eurac.edu/repository/xmlui/handle/20.500.12124/64
- Kolipsi 2 (V 1.1.): https://clarin.eurac.edu/repository/xmlui/handle/20.500.12124/66
- LEONIDE (V 1.1.): https://clarin.eurac.edu/repository/xmlui/handle/20.500.12124/25
Then, place XML files in corpora/ directory according to structure and run the extraction pipeline.
This work was conducted at EURAC Research, Institute for Applied Linguistics (November 2025 - January 2026) as part of a Language Technology Internship.
Supervision: Dr. Jennifer-Carmen Frey, Dr. Aivars Glaznieks (EURAC Research Institute for Applied Linguistics)
Corpus Providers:
- LEONIDE: Glaznieks et al. (2022)
- Kolipsi: Glaznieks et al. (2023)
Method and Evaluation Framework: Based on MultiLexNorm shared task (van der Goot et al., 2021) and the work described by Krupaninen et al. (2023)
Author: Lucia Galiero
GitHub: @lgaliero
Repository: LearnTextNorm-De
For questions or issues, please open a GitHub issue or contact via the repository.
Last Updated: February 11th, 2026