softcatala/Europarl-catalan

Dataset Card for Tilde-MODEL-Catalan

Dataset Summary

This dataset contains two dataset pairs corresponding to the Europarl corpus. Both the English and the German version are aligned with the Catalan translation, which has been obtained using Apertium's RBMT system from the Spanish version of the Spanish-English alignment. Catalan-German alignment has been obtained using this alignment finder from de-en and ca-en.

• Catalan-English: 1 965 735 segments.
• Catalan-German: 1 734 644 segments.

Supported Tasks and Leaderboards

This dataset can be used to train NMT and SMT systems. It has been used as a training corpus for the Softcatalà machine translation engine.

Languages

Catalan (ca). German (de). English (en).

Dataset Structure

Data Instances

Each instance is a line-aligned segment pair. For Catalan-English, the Catalan and English files are intended to be read in parallel. For Catalan-German, the Catalan and German files are intended to be read in parallel.

Data Fields

Raw text in compressed files:

• europarl.en-ca.ca.xz: Catalan side of the Catalan-English corpus.
• europarl.en-ca.en.xz: English side of the Catalan-English corpus.
• europarl.de-ca.ca.xz: Catalan side of the Catalan-German corpus.
• europarl.de-ca.de.xz: German side of the Catalan-German corpus.

Data Splits

The dataset provides training corpora for two language pairs, with one aligned file per language side.

Dataset Creation

Curation Rationale

The dataset was created to provide Catalan-English and Catalan-German parallel data for training statistical and neural machine translation systems.

Source Data

Initial Data Collection and Normalization

The corpus is based on Europarl v7. The Spanish original Europarl corpus was improved to fix spelling mistakes and other errors before producing the Catalan translation. Catalan sentences were translated from Spanish using Apertium rule-based machine translation. Catalan-German alignment was produced using a pivot-language alignment process from German-English and Catalan-English data.

Who are the source language producers?

The original source texts come from European Parliament proceedings in the Europarl corpus. The Catalan side was machine-generated and prepared by Softcatala.

Annotations

Annotation process

No manual linguistic annotation is added. The dataset processing consists of source cleanup, machine translation into Catalan, and sentence alignment.

Who are the annotators?

Not applicable.

Personal and Sensitive Information

The data is derived from public parliamentary proceedings. It may contain names of public figures or other references present in the original proceedings, but it was not created to collect personal or sensitive information.

Considerations for Using the Data

Social Impact of Dataset

The dataset supports Catalan machine translation research and development in parliamentary and formal-language domains.

Discussion of Biases

The corpus reflects European Parliament proceedings and therefore over-represents formal, institutional, political, and legislative language. It should not be treated as a broad sample of everyday Catalan.

Other Known Limitations

The Catalan side is machine-generated from Spanish and may contain translation errors or unnatural phrasing. Sentence alignment may also contain noise, especially for the Catalan-German pair produced through pivot alignment.

Additional Information

Dataset Curators

@softcatala @jordimas @davidcanovas

Licensing Information

CC BY 4.0.

Citation Information

If you use this dataset, cite the dataset repository:

Softcatala/Europarl-catalan: https://github.com/Softcatala/Europarl-catalan

Contributions

Softcatala, Jordi Mas, and David Canovas.