DistaLs: a Comprehensive Collection of Language Distance Measures

Rob van der Goot, Esther Ploeger, Verena Blaschke, Tanja SamardzicEMNLP 2025 · OpenReview

Abstract

Languages vary along a wide variety of dimensions. In Natural Language Processing (NLP), it is useful to know how “distant” languages are from each other, so that we can inform NLP models about these differences or predict good transfer languages. Furthermore, it can inform us about how diverse language samples are. However, there are many different perspectives on how distances across languages could be measured, and previous work has predominantly focused on either intuition or a single type of distance, like genealogical or typological distance. Therefore, we propose DistaLs, a toolkit that is designed to provide users with easy access to a wide variety of language distance measures. We also propose a filtered subset, which contains less redundant and more reliable features. DistaLs is designed to be accessible for a variety of use cases, and offers a Python, CLI, and web interface. It is easily updateable, and available as a pip package. Finally, we provide a case-study in which we use DistaLs to measure correlations of distance measures with performance on four different morphosyntactic tasks.

Introduction

In multilingual NLP, quantifying the similarity between languages is a critical task. This knowledge informs decisions on source language selection for transfer learning, performance prediction for unseen languages, and the assessment of dataset diversity. Historically, researchers have relied on fragmented tools or intuitive proxies (e.g., shared scripts), which can be misleading. For instance, cross-lingual transfer can succeed between seemingly distant languages like Indonesian and Irish. There is a clear need for a unified framework that integrates metadata, linguistic structure, and textual statistics.

The DistaLs Framework

DistaLs addresses this gap by aggregating a massive array of distance measures into a single, standardized interface. The toolkit harmonizes disparate data sources using the ISO 639-3 standard, ensuring all metrics are bi-directional and normalized to a 0.0 (identical) to 1.0 (maximal distance) scale.

Multidimensional Metrics

The toolkit categorizes distances into four primary domains:

  1. Metadata: Resource-based metrics such as Wikipedia size, speaker counts, and NLP resource availability.
  2. Typology: Structural features including phoneme inventories (PHOIBLE) and morphosyntax (Grambank).
  3. Wordlists: Lexical similarity based on databases like ASJP.
  4. Text-driven: Statistical measures derived from raw text, such as punctuation usage and character distributions.
Overview of features, sources, and language coverage in DistaLs.

DistaLs is designed for high accessibility, offering a Python package, a CLI, and a web-based Streamlit application. Crucially, it includes scraping scripts to allow the underlying database to be updated from original sources, preventing the tool from becoming a static, outdated dataset.

Analysis and Key Insights

A central finding of this work is that language distance is not monolithic. Languages may be genealogically close yet distant in terms of digital resources. DistaLs allows researchers to select the dimension most relevant to their specific task.

Redundancy and the Filtered Subset

Through correlation analysis, the authors identified significant redundancies among various distance measures. For example, a language's script serves as a strong proxy for many text-based distances. To simplify analysis for users, DistaLs proposes a "filtered subset" of reliable, non-overlapping features.

Correlation heatmap showing the relationships between different language distance features.

Predictive Power

In a case study involving Universal Dependencies (UD) v2.15 across 126 target languages, the authors found that nlp_state (resource availability) and grambank (structural typology) were the strongest predictors of performance on syntactic tasks, with Pearson correlations between -0.42 and -0.45. Interestingly, distance measures proved more sensitive to syntactic tasks (like dependency parsing) than to morphological ones (like lemmatization).

Conclusion and Future Work

DistaLs provides unprecedented coverage, supporting up to 7,856 languages for specific features. While it offers a robust framework, it is subject to the inherent biases of the ISO 639-3 standard and the sparsity of data for low-resource languages. Furthermore, the current text-driven metrics rely heavily on religious texts (Bible translations), which may not reflect modern usage. Future iterations will aim to incorporate more diverse corpora and address the monolithic treatment of languages by accounting for dialectal variation.

Summary updated

Abstract

Languages vary along a wide variety of dimensions. In Natural Language Processing (NLP), it is useful to know how “distant” languages are from each other, so that we can inform NLP models about these differences or predict good transfer languages. Furthermore, it can inform us about how diverse language samples are. However, there are many different perspectives on how distances across languages could be measured, and previous work has predominantly focused on either intuition or a single type of distance, like genealogical or typological distance. Therefore, we propose DistaLs, a toolkit that is designed to provide users with easy access to a wide variety of language distance measures. We also propose a filtered subset, which contains less redundant and more reliable features. DistaLs is designed to be accessible for a variety of use cases, and offers a Python, CLI, and web interface. It is easily updateable, and available as a pip package. Finally, we provide a case-study in which we use DistaLs to measure correlations of distance measures with performance on four different morphosyntactic tasks.
BibTeX
@inproceedings{goot2025distals,
  author = {Rob van der Goot and Esther Ploeger and Verena Blaschke and Tanja Samardzic},
  title = {{DistaLs: a Comprehensive Collection of Language Distance Measures}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=neyh9BD8cQ},
}
Figure