Czech Grammar Error Correction Corpus (GECCC)

The Czech Grammar Error Correction Corpus (GECCC) is a curated collection of Czech-language sentences that contain grammatical errors together with corrected versions of those sentences. It was created to support research and development of automatic grammarerror detection and correction (GED) systems for Czech, a language that historically received far less attention than English in this research area. The corpus is freely available for academic use and has already become a reference point for many studies in computational linguistics, language education, and natural language processing (NLP).

Why a Czech GED Corpus Matters

Czech is a highly inflected Slavic language with a rich system of case markings, verb aspect, gender agreement, and free word order. These properties make error detection challenging but also fascinating from a linguistic perspective. A dedicated corpus enables researchers to:

• Train supervised machinelearning models that understand Czech morphology.
• Evaluate the strengths and weaknesses of rulebased versus neural approaches.
• Develop tools for language learners, teachers, and editors.
• Compare error patterns across languages and discover typologically specific phenomena.

Corpus Construction

Data Sources

GECCC draws from several authentic sources:

• Student essays written for Czech language courses at secondary and university level.
• Online forums and social media where native speakers post informal text.
• Official documents that have been manually proofread, providing examples of typical professional errors.

Annotation Process

Each sentence was annotated by at least two nativespeaker linguists with expertise in Czech grammar. The workflow consisted of:

1. Error identification annotators marked the exact token(s) that were incorrect.
2. Error classification each error received a label from a predefined taxonomy (e.g., case agreement, verb aspect, preposition selection, punctuation).
3. Correction a corrected version of the sentence was written, preserving the original meaning as closely as possible.
4. Quality control a third annotator resolved any disagreements, and a final consistency check was performed automatically.

Taxonomy of Errors

The error taxonomy contains 12 major categories and 48 subcategories. Some of the most frequent types are:

• Case errors wrong case on nouns, adjectives, or pronouns.
• Agreement errors mismatches in gender, number, or person between subjects, verbs, and modifiers.
• Verb aspect using perfective instead of imperfective (or viceversa) where the context demands a specific aspect.
• Preposition choice selecting a preposition that governs an incorrect case.
• Word order placing elements in a sequence that violates Czech syntactic conventions.

Corpus Statistics (as of 2024)

Access and Licensing

The corpus is distributed under a CC BYNCSA 4.0 license. Researchers can download the data in two main formats:

• Plaintext (UTF8) each line contains a source sentence, a tab character, and the corrected sentence.
• CoNLLU a tokenlevel format that includes partofspeech tags, morphological features, error spans, and correction tags.

Documentation with detailed schema definitions, example files, and scripts for evaluation are provided on the projects GitHub repository.

Evaluation Benchmarks

Since its release, GECCC has been used to benchmark a wide range of GED systems. The most common evaluation metrics are:

• Precision, Recall, F0.5 favouring precision because false corrections are more detrimental to learners.
• GLEU an adaptation of BLEU that measures overlap between corrected and reference sentences while accounting for error spans.
• M2 scorer the standard metric used in the CoNLL2014 and BEA2019 shared tasks.

Key Research Findings Using GECCC

1. Neural seq2seq models achieve higher recall than traditional rulebased tools, but often overcorrect, leading to lower precision (average F0.5 0.42).
2. Morphologyaware transformers that incorporate Czech lemmatization and POS tags improve both precision and recall (F0.5 up to 0.56).
3. Curriculum learning training first on synthetic errors, then finetuning on GECCC, yields the best tradeoff for lowresource scenarios.
4. Errortype analysis shows that case and agreement errors remain the hardest for models, whereas punctuation and wordorder errors are corrected more reliably.
5. Crosslingual transfer models pretrained on larger English GED corpora can be adapted to Czech with modest performance loss, highlighting the value of multilingual pretraining.

Tools Built on GECCC

Several publicly available applications rely on the corpus for training or evaluation:

• CZGEC an opensource grammarchecking plugin for LibreOffice.
• LangCorrector a webbased writing assistant that highlights errors and suggests corrections in real time.
• Educational dashboards used by Czech language departments to track common learner mistakes.

How to Contribute

The community is encouraged to expand GECCC in three ways:

• Submit new sentences from underrepresented registers (e.g., spoken dialogues, technical manuals).
• Annotate additional error types such as style or lexical choice, extending the current taxonomy.
• Provide baseline systems and sharedtask results to keep the benchmark uptodate.

All contributions are merged after a doubleblind review process to preserve annotation quality.

Future Directions

Looking ahead, the GECCC project aims to:

1. Increase the size of the corpus to 100k sentences, with a focus on learner data from primary and secondary schools.
2. Add audio alignments linking spoken utterances to corrected transcripts, enabling speechtotext GED research.
3. Integrate semantic error detection, covering meaningpreserving slips such as collocation misuse.
4. Release a multilingual errortype mapping that aligns Czech error categories with those in other Slavic languages, facilitating crosslingual studies.

Getting Started

To experiment with the corpus, follow these simple steps:

1. Clone the repository: git clone https://github.com/geccc/czech-gec.git
2. Install the required Python packages (shown in requirements.txt).
3. Run the provided dataloader script to obtain training, validation, and test splits.
4. Train a baseline transformer model using the train.py script; a readymade configuration for Hugging Face models is included.
5. Evaluate using the m2scorer.py script to obtain precision, recall, and F0.5 scores.

For detailed tutorials and example notebooks, see the examples directory.

References: LREC 2022 Introducing GECCC, ACL 2023 Demo Morphologyaware Transformers for Czech GEC.