The ITU Faroese Pairs Dataset

Leon Derczynski, Annika Solveig Hedegaard Isfeldt, Signhild DjurhuusarXiv 2022 · OpenReview

Abstract

This article documents a dataset of sentence pairs between Faroese and Danish, produced at ITU Copenhagen. The data covers translation from both source languages and is intended for use as training data for machine translation systems in this language pair.

Introduction

In the landscape of Natural Language Processing (NLP), linguistic representation is highly skewed. While high-resource languages like English and Mandarin Chinese benefit from massive datasets, many others remain "digitally invisible." Faroese, a North Germanic language spoken by approximately 70,000 people, is a prime example. Under the classification framework proposed by Joshi et al. (2020), Faroese is categorized as a "Class 1" language—one with extremely limited digital resources. This lack of data hinders the development of modern tools like machine translation (MT), potentially isolating the language from global technological integration and threatening its cultural preservation.

To address this gap, researchers at ITU Copenhagen have developed the ITU Faroese Pairs Dataset. This resource provides approximately 4,000 human-translated sentence pairs between Faroese and Danish, offering a high-quality parallel corpus designed to kickstart MT development for this underserved language pair.

Methodology and Data Documentation

The creation of the ITU Faroese Pairs Dataset was guided by the "Data Statement for NLP" framework (Bender and Friedman, 2018). This approach emphasizes transparency in dataset construction, documenting the demographics and motivations of the translators to provide context for future researchers.

Translation Process

The dataset was produced using a human-in-the-loop methodology involving two native (L1) Faroese speakers. Both translators were highly proficient in Danish and resided in Denmark at the time of the project, ensuring a deep understanding of the nuances in both the source and target languages.

The translation guidelines prioritized linguistic naturalness over rigid literalism. Annotators were instructed to produce fluent, idiomatic sentences in the target language. This focus on "fluency over literalism" is critical for training neural machine translation (NMT) models that can generate human-like text rather than clunky, word-for-word substitutions. The workflow was managed asynchronously, with translators processing blocks of 100–200 sentences via shared spreadsheets.

Dataset Composition

The ITU Faroese Pairs Dataset is curated from three distinct sources to capture a variety of linguistic "registers" and styles:

  1. Dimmalætting: This subset includes sentences from a prominent Faroese newspaper (2015–2019). It captures formal news styles, regional names, and contemporary cultural concepts.
  2. Europarl: Drawn from European Parliament proceedings, this source provides formal, political, and legal terminology.
  3. Tatoeba: To balance the formal sources, Tatoeba provides informal, everyday conversational sentences, ensuring the dataset reflects how people speak in daily life.

The final collection comprises roughly 4,000 sentence pairs. The dataset is released under a CC-BY 4.0 license and is hosted on the Hugging Face Datasets platform for easy integration into modern NLP pipelines.

Key Insights and Strategic Design

A central contribution of this work is the strategic mixing of registers. By combining news, political discourse, and informal chat, the authors ensure the dataset is not a monolith. This diversity is essential for training robust models capable of handling various contexts, from official reports to casual communication.

Furthermore, the project highlights the "Class 1" digital divide. While Danish is a "Class 2" language (reasonably well-resourced), its proximity to Faroese makes it an ideal bridge. However, the researchers note that other North Atlantic languages, such as Greenlandic and certain Danish dialects like Sønderjysk, remain similarly neglected.

The reliance on L1 translators living in the target culture (Denmark) was a deliberate choice. For a population that is often bilingual by necessity, capturing the subtle interplay between Faroese and Danish requires translators who are immersed in both linguistic environments.

Results and Impact

The primary outcome of this research is the establishment of a high-quality benchmark for Faroese-Danish machine translation. By making the data accessible via Hugging Face, the researchers have lowered the barrier to entry for other teams to experiment with Faroese NLP.

Additionally, the project sets a standard for documentation in low-resource language research. By providing clear demographic information and following established data statement protocols, the authors provide a template for how small-scale, high-quality datasets should be presented to the scientific community.

Limitations and Future Directions

Despite its significance, the dataset has inherent limitations:

  • Scale: While 4,000 pairs represent a major milestone for Faroese, modern deep learning models typically require millions of examples to achieve high-level fluency.
  • Subjectivity: With only two translators, the dataset reflects their specific idiolects and stylistic preferences.
  • Quality Assurance: The project relied on the expertise of the native speakers without a secondary blind review process or automated quality metrics.

Future work will focus on using this dataset as a "seed" for fine-tuning larger multilingual models. By leveraging transfer learning from high-resource languages, researchers hope to bridge the performance gap and finally bring Faroese into the fold of modern translation technology.

Summary updated

Abstract

This article documents a dataset of sentence pairs between Faroese and Danish, produced at ITU Copenhagen. The data covers tranlsation from both source languages, and is intended for use as training data for machine translation systems in this language pair.
BibTeX
@misc{derczynski2022itu,
  author = {Leon Derczynski and Annika Solveig Hedegaard Isfeldt and Signhild Djurhuus},
  title = {{The ITU Faroese Pairs Dataset}},
  year = {2022},
  eprint = {2206.08727},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.08727},
}
Figure