Data Augmentation for Low-Resource Languages

98 concepts88 papers2 proposals

The performance of modern Natural Language Processing (NLP) models scales dramatically with data size. However, for the vast majority of the world's languages, dialects, and historical variants, large annotated datasets simply do not exist. As noted in the Findings of the IWSLT 2025 Evaluation Campaign, standard supervised learning approaches for languages with minimal training data have reached a plateau. To break through this performance ceiling, researchers rely on data augmentation—the artificial creation of training examples to teach models linguistic patterns when natural data is scarce.

The Spectrum of Augmentation: From Naive Noise to Linguistic Rules

How do we create new text without human annotators? The simplest approach is to apply surface-level perturbations that do not require external linguistic databases. For instance, in AEDA: An Easier Data Augmentation Technique for Text Classification, Akbar Karimi 0001 et al. demonstrate that randomly inserting punctuation marks into training sentences can significantly boost classification performance. Because this technique preserves the original word order and prevents information loss, it functions as an effective regularizer without needing complex language models.

However, while naive methods are computationally cheap, they often produce ungrammatical strings that fail to capture complex morphological structures. This is a major hurdle for morphologically rich low-resource languages. To address this, linguistically-motivated approaches inject expert knowledge directly into the generation pipeline. Toms Bergmanis et al. demonstrate this in Training Data Augmentation for Context-Sensitive Neural Lemmatizer Using Inflection Tables and Raw Text by pairing rule-based morphological inflection tables with raw, unannotated text. This hybrid approach expands the vocabulary of neural models and dramatically improves their performance on unseen words. Similarly, Jenna Kanerva et al. use rule-based tools alongside autoencoding tasks in Universal Lemmatizer: A sequence-to-sequence model for lemmatizing Universal Dependencies treebanks to teach models identity mappings and morphological patterns simultaneously.

The choice between these two paradigms involves a fundamental trade-off. As explored by Ray Groshan et al. in Is linguistically-motivated data augmentation worth it?, the ultimate success of an augmentation strategy depends heavily on how closely the synthetic data mimics the natural distribution of the target language. While naive methods provide cheap regularization, linguistically-informed transformations offer a critical edge when structural correctness is paramount.

Cross-Lingual Projection and Structural Preservation

When a target language lacks resources entirely, another powerful strategy is to borrow data from a high-resource counterpart. This can be achieved via machine translation or back-translation, as shown by Anindo Barua et al. in CUET_Absolute_Zero@DravidianLangTech 2025: Detecting AI-Generated Product Reviews in Malayalam and Tamil Language Using Transformer Models. By translating existing corpora into low-resource Dravidian languages, the authors generated sufficient training data to robustly train classification models.

However, a major challenge arises when the target task depends on complex structural annotations, such as syntactic trees or semantic labels. Standard translation pipelines often break these structures. To solve this, Jayeol Chun et al. in Modal Dependency Parsing via Biaffine Attention with Self-Loop leverage large language models to translate text while explicitly preserving structural annotations. This cross-lingual projection allows specialized linguistic parsers to be trained in low-resource environments without requiring manual re-annotation of the translated text.

Generative Bootstrapping and Quality Filtering

With the advent of Large Language Models (LLMs), researchers can now prompt generative models to synthesize entirely new training corpora from scratch. For example, in Visibility as Survival: Generalizing NLP for Native Alaskan Language Identification, Ivory Yang et al. introduce a framework that uses generative models to expand extremely limited datasets for endangered Native Alaskan languages, enabling near-perfect language identification.

Yet, purely generative data introduces a new risk: model hallucinations and unnatural phrasing. In a comparative study of code-switched machine translation, Injy Hamed et al. discovered a striking 0.97 correlation between the "naturalness" of synthetic data and its actual effectiveness in training downstream models (Data Augmentation Techniques for Machine Translation of Code-Switched Texts: A Comparative Study).

To bridge the gap between purely synthetic generation and high-quality natural text, modern frameworks employ "model-in-the-loop" filtering and validation. In A Simple Yet Effective Corpus Construction Framework for Indonesian Grammatical Error Correction, Meiyu Zeng et al. demonstrate a pipeline that combines synthetic pre-training with automated error discovery and human-verified refinement. Similarly, Nils Dycke et al. show in Overview of PragTag-2023: Low-Resource Multi-Domain Pragmatic Tagging of Peer Reviews that filtering synonym-replaced synthetic examples using semantic similarity metrics ensures the augmented data remains linguistically coherent. By combining generative models with strict validation metrics, researchers can scale up training sets while maintaining the distributional integrity required for robust model performance.

Key Foundational Work

Proceedings of the First Workshop on Multilingual Surface Realisation (2018) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_data_augmentation_for_low_resource_langu_d0b06cc9,
  title = {data augmentation for low-resource languages},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-augmentation-for-low-resource-languages/rdr_d0b06cc92b4b415c84fb4980bb6a3785},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_data_augmentation_for_low_resource_langu_d0b06cc9_v1,
  title = {data augmentation for low-resource languages},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-augmentation-for-low-resource-languages/rdr_d0b06cc92b4b415c84fb4980bb6a3785/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake