README.md

metadata

license: cc-by-4.0
task_categories:
  - translation
language:
  - gl
  - es
  - en
size_categories:
  - 100K<n<1M

Dataset Card for Spanish–Galician / English–Galician Scientific Corpus (SciELO)

Dataset Summary

The SciELO Corpus, hosted in theOPUS repository, is a large-scale parallel resource composed of full scientific articles extracted from the Scientific Electronic Library Online (SciELO). It provides high-quality sentence pairs between Spanish, Portuguese, and English across diverse academic domains.

This corpus is particularly valuable for training machine translation models, as it offers specialized terminology density and complex grammatical structures that reflect real scientific and technical language usage in Latin America and Iberia.

To address the lack of publicly available Galician data in this domain, Portuguese segments were adapted into Galician using transliteration and localization tools found in our text pipeline and Apertium. The resulting texts were then normalized through our cleaning pipeline, ensuring consistency and readiness for model development.

The final resource is a parallel scientific corpus of ~300,000 aligned sentences for the pairs Spanish–Galician and English–Galician.

Dataset Creation

• Apertium pt-gl: We translate original Portuguese segments into Galician using symbolic rules.
• Transliteration and localization: Using port2gal, we improve Apertium's output by processing all leftover tags for out-of-vocabulary words, which we either transliterate to Spanish orthography or localize to a more common Galician word.
• Encoding errors: The entire text was scanned for encoding errors ensuring it is utf-8 encoded.
• Deduplication: The filtered datasets were deduplicated to remove redundant sentence pairs.
• Pyplexity: We used pyplexity to filter texts that may contain non-linguistic content.
• Normalization: The final Galician text present here was normalized (linguistically) to adhere to the autonomous standard of the Galician language.

Supported Tasks and Benchmarks

• Machine Translation (MT): Training and evaluation of MT systems for Galician in scientific domains.
• Terminology Extraction: Building specialized bilingual glossaries for scientific and technical fields.
• Cross-lingual NLP: Supporting multilingual embeddings and semantic alignment in academic texts.
• Evaluation: Benchmarking translation quality in specialized scientific corpora.

Languages

• Spanish-Galician, English-Galician

Dataset Structure

• Format: Parallel text segments (aligned sentences)
• Domain: Scientific articles (multiple disciplines)
• Size: ~300,000 lines

Use Cases

• Training MT models for Galician in specialized scientific contexts.
• Creating bilingual dictionaries and glossaries for academic terminology.
• Supporting research in cross-lingual transfer for underrepresented languages.
• Evaluating translation systems in scientific and technical domains.

Limitations

• Domain-specific: Focused on scientific texts, which may not generalize to everyday language.
• Automatically adapted Galician segments may contain transliteration or translation artifacts.

Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA. Esta publicación del proyecto Desarrollo de Modelos ALIA está financiada por el Ministerio para la Transformación Digital y de la Función Pública y por el Plan de Recuperación, Transformación y Resiliencia – Financiado por la Unión Europea – NextGenerationEU.