Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NSL-MT: Linguistically Informed Negative Samples for Efficient Machine Translation in Low-Resource Languages

Mamadou K. Keita, Christopher Homan, Huy Le

TL;DR

NSL-MT tackles low-resource machine translation by injecting linguistically informed negative evidence into the training objective, explicitly penalizing outputs that violate target-language grammar. By generating hard negatives across morphological, syntactic, and lexical dimensions with severity weights, NSL-MT complements standard likelihood training and yields large gains in data-constrained scenarios while improving cross-architecture performance. The method achieves up to 89% BLEU improvements on weaker baselines and demonstrates a practical 5x data-efficiency benefit, especially at 1,000 examples, by guiding models toward correct target-language structure with explicit negative signals. Across languages and architectures, NSL-MT reduces form-related errors and preserves meaning, offering a scalable approach to enhance translation quality with limited parallel data and minimal additional linguistic resources.

Abstract

We introduce Negative Space Learning MT (NSL-MT), a training method that teaches models what not to generate by encoding linguistic constraints as severity-weighted penalties in the loss function. NSL-MT increases limited parallel data with synthetically generated violations of target language grammar, explicitly penalizing the model when it assigns high probability to these linguistically invalid outputs. We demonstrate that NSL-MT delivers improvements across all architectures: 3-12\% BLEU gains for well-performing models and 56-89\% gains for models lacking descent initial support. Furthermore, NSL-MT provides a 5x data efficiency multiplier -- training with 1,000 examples matches or exceeds normal training with 5,000 examples. Thus, NSL-MT provides a data-efficient alternative training method for settings where there is limited annotated parallel corporas.

NSL-MT: Linguistically Informed Negative Samples for Efficient Machine Translation in Low-Resource Languages

TL;DR

NSL-MT tackles low-resource machine translation by injecting linguistically informed negative evidence into the training objective, explicitly penalizing outputs that violate target-language grammar. By generating hard negatives across morphological, syntactic, and lexical dimensions with severity weights, NSL-MT complements standard likelihood training and yields large gains in data-constrained scenarios while improving cross-architecture performance. The method achieves up to 89% BLEU improvements on weaker baselines and demonstrates a practical 5x data-efficiency benefit, especially at 1,000 examples, by guiding models toward correct target-language structure with explicit negative signals. Across languages and architectures, NSL-MT reduces form-related errors and preserves meaning, offering a scalable approach to enhance translation quality with limited parallel data and minimal additional linguistic resources.

Abstract

We introduce Negative Space Learning MT (NSL-MT), a training method that teaches models what not to generate by encoding linguistic constraints as severity-weighted penalties in the loss function. NSL-MT increases limited parallel data with synthetically generated violations of target language grammar, explicitly penalizing the model when it assigns high probability to these linguistically invalid outputs. We demonstrate that NSL-MT delivers improvements across all architectures: 3-12\% BLEU gains for well-performing models and 56-89\% gains for models lacking descent initial support. Furthermore, NSL-MT provides a 5x data efficiency multiplier -- training with 1,000 examples matches or exceeds normal training with 5,000 examples. Thus, NSL-MT provides a data-efficient alternative training method for settings where there is limited annotated parallel corporas.

Paper Structure

This paper contains 35 sections, 4 equations, 1 figure, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Negative Space Learning
  3. Core Principle
  4. Violation Generation
  5. Training Objective
  6. Implementation
  7. Experiments and Results
  8. Experimental Setup
  9. Datasets
  10. Models
  11. Training Configuration
  12. Evaluation Metrics
  13. Results
  14. Ablation Study
  15. Data Efficiency Analysis
  16. ...and 20 more sections

Figures (1)

  • Figure 1: Data efficiency comparison between Normal training(in red) and NSL-MT(in green) across varying training set sizes. NSL-MT achieves high performance at all data sizes, with the largest relative gains occurring at the smallest data sizes.