Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Morphologically-Aware Dictionary-based Data Augmentation Technique for Machine Translation of Under-Represented Languages

Md Mahfuz Ibn Alam, Sina Ahmadi, Antonios Anastasopoulos

TL;DR

This work tackles data scarcity in machine translation for under-represented languages by introducing a morphology-aware, dictionary-based data augmentation method that uses a seed parallel corpus and PanLex bilingual lexicons. The approach builds synthetic parallel data through a four-step pipeline—alignment, analysis, morphologically informed replacement, and generation—then filters augmented sentences with a language model before training MT systems, either from scratch or by fine-tuning DeltaLM. Across 14 English to X language pairs, the morphologically informed augmentation yields consistent BLEU improvements, often with only 5K synthetic examples, highlighting the value of morphology preservation in synthetic data. The method is particularly effective for low-resource languages and can complement other augmentation techniques, offering practical gains where parallel data remain scarce.

Abstract

The availability of parallel texts is crucial to the performance of machine translation models. However, most of the world's languages face the predominant challenge of data scarcity. In this paper, we propose strategies to synthesize parallel data relying on morpho-syntactic information and using bilingual lexicons along with a small amount of seed parallel data. Our methodology adheres to a realistic scenario backed by the small parallel seed data. It is linguistically informed, as it aims to create augmented data that is more likely to be grammatically correct. We analyze how our synthetic data can be combined with raw parallel data and demonstrate a consistent improvement in performance in our experiments on 14 languages (28 English <-> X pairs) ranging from well- to very low-resource ones. Our method leads to improvements even when using only five seed sentences and a bilingual lexicon.

A Morphologically-Aware Dictionary-based Data Augmentation Technique for Machine Translation of Under-Represented Languages

TL;DR

This work tackles data scarcity in machine translation for under-represented languages by introducing a morphology-aware, dictionary-based data augmentation method that uses a seed parallel corpus and PanLex bilingual lexicons. The approach builds synthetic parallel data through a four-step pipeline—alignment, analysis, morphologically informed replacement, and generation—then filters augmented sentences with a language model before training MT systems, either from scratch or by fine-tuning DeltaLM. Across 14 English to X language pairs, the morphologically informed augmentation yields consistent BLEU improvements, often with only 5K synthetic examples, highlighting the value of morphology preservation in synthetic data. The method is particularly effective for low-resource languages and can complement other augmentation techniques, offering practical gains where parallel data remain scarce.

Abstract

The availability of parallel texts is crucial to the performance of machine translation models. However, most of the world's languages face the predominant challenge of data scarcity. In this paper, we propose strategies to synthesize parallel data relying on morpho-syntactic information and using bilingual lexicons along with a small amount of seed parallel data. Our methodology adheres to a realistic scenario backed by the small parallel seed data. It is linguistically informed, as it aims to create augmented data that is more likely to be grammatically correct. We analyze how our synthetic data can be combined with raw parallel data and demonstrate a consistent improvement in performance in our experiments on 14 languages (28 English <-> X pairs) ranging from well- to very low-resource ones. Our method leads to improvements even when using only five seed sentences and a bilingual lexicon.
Paper Structure (34 sections, 1 equation, 3 figures, 6 tables)

This paper contains 34 sections, 1 equation, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Analysis
  4. Alignment
  5. Morphological Tagging
  6. Word-pair Selection
  7. Augmentation
  8. Morphologically-Informed
  9. Naive
  10. Filtering with LMs
  11. Neural Machine Translation
  12. Experimental Setup
  13. Dataset
  14. Parallel Data
  15. Bilingual Dictionaries
  16. ...and 19 more sections

Figures (3)

  • Figure 1: A schema of our approach. After aligning 'guitar' (in English) and 'gîtarê' (in Kurmanji Kurdish), the new word 'flower' is randomly selected to replace 'guitar' and its translation 'gul' in a bilingual dictionary is inflected according to its morphological features as 'gulê'. Small caps refer to lemmata.
  • Figure 2: BLEU scores on the test sets for six languages in the english-x direction. X-axis indicates the amount of synthetic parallel data we use along with seed data. The baseline uses no synthetic data. Except for Irish and Galician, all the other languages do not benefit from the increasing amounts of synthetic data. It seems like Irish has even room for more improvement. ours is the morphologically-informed method.
  • Figure 3: BLEU scores on the test sets for six languages in the x-English direction. ours is the morphologically-informed method. Morphologically-informed approach outperforms the naive approach in all the six language pairs. X-axis indicates the amount of synthetic parallel data we use along with seed data. The baseline uses no synthetic data.