oliat/lemmatization

Dataset Details

Dataset Description

This dataset contains lemmatization data for seven languages: Basque, Czech, English, Polish, Russian, Spanish, and Turkish. For each language, the data is provided in three different Shortest Edit Script (SES) representations:

• ses-udpipe
• ses-ixapipes
• ses-morpheus

The datasets were created for the experiments presented in Evaluating Shortest Edit Script Methods for Contextual Lemmatization (Toporkov and Agerri (2024)) and are intended for multilingual contextual lemmatization. The files are provided in a format ready for model training and can be used to fine-tune pretrained language models such as XLM-RoBERTa-large as a token classification task, where the model predicts an edit script that is subsequently applied to generate the lemma. The three SES variants encode the transformation from a surface word form to its lemma using different edit-script generation strategies, allowing the effect of lemma representation on contextual lemmatization performance to be studied across languages with different morphological profiles.

• Curated by: Olia Toporkov and Rodrigo Agerri
• Language(s): Basque (eu), Czech (cs), English (en), Polish (pl), Russian (ru), Spanish (es), and Turkish (tr)
• License: Apache-2.0

Dataset Sources

• Paper: Evaluating Shortest Edit Script Methods for Contextual Lemmatization
• Original code and resources: hitz-zentroa/ses-lemma
• Project: Multilingual Contextual Lemmatization

Uses

Direct Use

The dataset is intended for training and evaluating contextual lemmatization models that formulate lemmatization as a token classification task.

In particular, it can be used to:

• train models to predict Shortest Edit Script labels from words in context;
• compare different SES representations for contextual lemmatization;
• evaluate multilingual and language-specific pretrained language models;
• study lemmatization across languages with different morphological characteristics;
• reproduce or extend the experiments described in Toporkov and Agerri (2024).

The data can be used with multilingual pretrained encoders such as XLM-RoBERTa, as well as suitable language-specific encoder models.

Dataset Structure

The dataset is organized by language and Shortest Edit Script representation. Each SES variant contains the corresponding data used for training and evaluation. Depending on the language and experimental setting, the resources include training, development, in-domain test, and out-of-domain test data. The SES label associated with each token encodes the sequence of transformations required to convert the observed word form into its lemma.

For each of the seven languages, three SES variants are provided along with all-labels files listing the unique SES labels found in the training, development, and test corpora, and requiered to reproduce the experiments:

• ses-udpipe
• ses-ixapipes
• ses-morpheus

The structure of the files in each ses-folder is the following:

• in-domain-train-file
• in-domain-dev-file
• in-domain-test-file
• out-of-domain-test-file
• all-labels-file

Source Data

The experimental data originates from existing linguistically annotated corpora and treebanks used in the corresponding publication.

The in-domain datasets include:

• Spanish: GSD
• Russian: GSD
• English: English Web Treebank (EWT)
• Basque: Basque Dependency Treebank (BDT)
• Turkish: ITU-METU-Sabancı Treebank (IMST)
• Czech: Czech Academic Corpus (CAC)
• Polish: LFG corpus

The paper additionally evaluates the models on out-of-domain corpora, including resources such as AnCora, SynTagRus, GUM, Armiarma, and PUD, depending on the language. Please refer to the paper for the complete description and citations of the original corpora.

Data Collection and Processing

The original corpora provide tokenized text together with lemma annotations. For every word–lemma pair, three alternative SES representations were automatically generated:

1. UDPipe SES (ses-udpipe)
   Based on the lemmatization approach used in UDPipe (Straka et al., 2016). It encodes casing information separately from character-level edit operations and uses the longest common substring between the word form and lemma to identify the unchanged part.

2. IXA pipes SES (ses-ixapipes)
   Based on the edit-script approach used in IXA pipes (Agerri et al., 2014), which in turn builds on the lemma-class representation introduced by Chrupała et al. (2008).

3. Morpheus SES (ses-morpheus)
   Based on Morpheus (Yildiz and Tantuğ, 2019), which represents the transformation from a surface form to its lemma through character-level edit operations such as keeping, deleting, replacing, and inserting characters.

These SES labels are used as classification labels during model fine-tuning.

Annotations

The lemma annotations originate from the corresponding source corpora. The Shortest Edit Script labels provided in this dataset are automatically derived from the word-form and lemma pairs using the three SES generation methods.

Recommendations

Users comparing SES representations should keep training data, model architecture, preprocessing, and evaluation settings consistent whenever possible. For full reproducibility of the experiments, please consult the original paper and accompanying code repository.

Citation

If you use this dataset, please cite:

Toporkov, Olia and Rodrigo Agerri. 2024. “Evaluating Shortest Edit Script Methods for Contextual Lemmatization.” In Proceedings of LREC-COLING 2024, pp. 6451–6463.

BibTeX

@inproceedings{toporkov-agerri-2024-evaluating,
    title = "Evaluating Shortest Edit Script Methods for Contextual Lemmatization",
    author = "Toporkov, Olia  and
      Agerri, Rodrigo",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.572/",
    pages = "6451--6463"
}

More Information

For additional information about the SES implementations, experimental settings, model configurations, and evaluation results, see:

• Paper
• Code repository
• Multilingual Contextual Lemmatization project page

Dataset Card Contact

For questions, bug reports, or issues related to the dataset, please open an issue in the associated GitHub repository.