Leveraging Loanword Constraints for Improving Machine Translation in a Low-Resource Multilingual Context

Felermino D. M. A. Ali, Henrique Lopes Cardoso, Rui Sousa-SilvaEMNLP 2025

TL;DR

This paper tackles the challenge of improving machine translation for low-resource languages that heavily borrow words from dominant languages. They propose a novel method of augmenting source sentences with explicit loanword constraints, which significantly boosts translation quality and accuracy in handling these borrowed terms.

What Problem Does This Paper Address?

Machine translation (MT) systems often struggle with low-resource languages, especially when these languages frequently incorporate words from dominant languages (lexical borrowing). This phenomenon is common in post-colonial contexts where a dominant language is used alongside indigenous ones. Human translators naturally adapt these borrowed words by applying phonetic rules or creating new terms. However, MT models lack this nuanced understanding, leading to inaccurate translations. The problem is further compounded by the lack of standardized orthography in some languages and the complex grammatical integration of borrowed words. This paper aims to bridge this gap by providing MT models with explicit knowledge of loanword correspondences.

Methods

The core of the proposed method involves loanword-constrained machine translation. This is achieved through data augmentation techniques applied to the source sentences. Two primary strategies are employed:

  • Replacement: Identified loanwords in the source sentence are replaced with a structured schema that explicitly links the source word to its target equivalent (e.g., <|start|> source_word <|translate-as|> target_word <|end|>). This method is inspired by existing work on encoder-decoder architectures.
  • Prompting: For decoder-only models (like Large Language Models - LLMs), the source sentence is augmented with constraints using a prompt-based format. This involves providing explicit instructions within the prompt to guide the model in translating loanwords correctly.

Both methods aim to enforce a "copy behavior" where the model learns to map source loanwords to their target counterparts. The authors also explored fine-tuning both multilingual Neural Machine Translation (NMT) models and various LLMs using these augmented datasets.

Datasets

The primary dataset used is a Portuguese-Emakhuwa bitext dataset (Ali et al., 2024a). This dataset includes parallel texts and manually annotated loanword constraints, particularly in the news subset. To address the scarcity of annotations, the researchers employed weak supervision using a fine-tuned GPT40-mini model to automatically annotate the remaining data. For evaluation, they utilized the FLORES+ DEV and DEVTEST datasets (Ali et al., 2024c), which are specifically designed for Portuguese-Emakhuwa MT evaluation and include manually annotated loanword information.

Key Insights and Ideas

  • Explicit loanword constraints are crucial: Simply providing parallel data isn't enough for low-resource languages with significant borrowing. Explicitly telling the model how to translate specific words makes a huge difference.
  • LLMs can be guided effectively: Even powerful LLMs, which can sometimes be unpredictable, show significant improvements when guided by loanword constraints through prompting.
  • Asymmetry in translation: The improvements from constraint-aware training are consistently greater in the Portuguese-to-Emakhuwa (pt→vmw) direction. This highlights the inherent challenges of translating into low-resource languages due to limited target-side vocabulary and data.
  • Robustness to input mismatch: Models trained on a mix of constrained and unconstrained inputs are more robust when faced with inputs that don't perfectly match their training conditions.

Key Results

The study demonstrates substantial improvements in translation quality when loanword constraints are applied:

  • Automatic Evaluation: Models trained with constraint injection consistently outperformed their unconstrained counterparts across various metrics (BLEU, CHRF, AfriCOMET). Notably, Sentence-Level Constraint Accuracy (SCA), a metric measuring how accurately loanwords are translated, saw significant gains, reaching up to 95% F1-score for the loanword constraint generation model itself.
  • Human Evaluation: Human evaluators found that constraint-trained models generally achieved higher adequacy scores compared to unconstrained models, with improvements ranging from 52% to 98%.
  • Model Performance: Both encoder-decoder (e.g., NLLB) and decoder-only (e.g., Llama, Qwen) architectures benefited from constraint training, with LLMs showing particularly impressive gains.

Limitations and Future Work

  • Data Scarcity: The experiments are limited to the Portuguese-Emakhuwa pair due to the scarcity of parallel corpora with explicit loanword annotations. Broader language coverage is needed to assess generalizability.
  • Evaluation Metrics: While AfriCOMET is used, its support for Emakhuwa is noted as potentially limited, suggesting future work on adapting or validating metrics for this specific context.
  • Constraint Integration Strategies: The study explored "replacement" and "prompting," but a more exhaustive exploration of alternative prompt formulations and schema variations is suggested.
  • Future Directions: The authors propose extending this method to incorporate bilingual dictionary entries for unseen words and balancing the use of culturally grounded terms with accurate loanword adaptation. They also suggest exploring how this approach can promote the use of native terminology.

Summary updated

BibTeX
@inproceedings{ali2025leveraging,
  author = {Felermino D. M. A. Ali and Henrique Lopes Cardoso and Rui Sousa-Silva},
  title = {{Leveraging Loanword Constraints for Improving Machine Translation in a Low-Resource Multilingual Context}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure