Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Priest to Doctor: Domain Adaptation for Low-Resource Neural Machine Translation

Ali Marashian, Enora Rice, Luke Gessler, Alexis Palmer, Katharina von der Wense

TL;DR

This paper tackles domain adaptation for neural machine translation under a realistic low-resource regime, using only Bible-domain parallel data, high-resource monolingual target-domain text, and a bilingual dictionary to translate from English into several low-resource languages. It compares four methods (DALI, LeCA, CPT, and a Combined approach) with mBART as a baseline across government and medical domains, finding that the simple DALI method most consistently boosts performance, though overall scores remain modest and human evaluation reveals room for improvement. The work highlights the limitations of current DA techniques in truly low-resource settings and suggests that leveraging in-domain monolingual data with lexicon-informed pseudo-parallel data (as in DALI) is a promising direction, while emphasizing the need for more robust methods. Code and data resources are made available to support replication and further research in this challenging area.

Abstract

Many of the world's languages have insufficient data to train high-performing general neural machine translation (NMT) models, let alone domain-specific models, and often the only available parallel data are small amounts of religious texts. Hence, domain adaptation (DA) is a crucial issue faced by contemporary NMT and has, so far, been underexplored for low-resource languages. In this paper, we evaluate a set of methods from both low-resource NMT and DA in a realistic setting, in which we aim to translate between a high-resource and a low-resource language with access to only: a) parallel Bible data, b) a bilingual dictionary, and c) a monolingual target-domain corpus in the high-resource language. Our results show that the effectiveness of the tested methods varies, with the simplest one, DALI, being most effective. We follow up with a small human evaluation of DALI, which shows that there is still a need for more careful investigation of how to accomplish DA for low-resource NMT.

From Priest to Doctor: Domain Adaptation for Low-Resource Neural Machine Translation

TL;DR

This paper tackles domain adaptation for neural machine translation under a realistic low-resource regime, using only Bible-domain parallel data, high-resource monolingual target-domain text, and a bilingual dictionary to translate from English into several low-resource languages. It compares four methods (DALI, LeCA, CPT, and a Combined approach) with mBART as a baseline across government and medical domains, finding that the simple DALI method most consistently boosts performance, though overall scores remain modest and human evaluation reveals room for improvement. The work highlights the limitations of current DA techniques in truly low-resource settings and suggests that leveraging in-domain monolingual data with lexicon-informed pseudo-parallel data (as in DALI) is a promising direction, while emphasizing the need for more robust methods. Code and data resources are made available to support replication and further research in this challenging area.

Abstract

Many of the world's languages have insufficient data to train high-performing general neural machine translation (NMT) models, let alone domain-specific models, and often the only available parallel data are small amounts of religious texts. Hence, domain adaptation (DA) is a crucial issue faced by contemporary NMT and has, so far, been underexplored for low-resource languages. In this paper, we evaluate a set of methods from both low-resource NMT and DA in a realistic setting, in which we aim to translate between a high-resource and a low-resource language with access to only: a) parallel Bible data, b) a bilingual dictionary, and c) a monolingual target-domain corpus in the high-resource language. Our results show that the effectiveness of the tested methods varies, with the simplest one, DALI, being most effective. We follow up with a small human evaluation of DALI, which shows that there is still a need for more careful investigation of how to accomplish DA for low-resource NMT.

Paper Structure

This paper contains 31 sections, 2 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Domain Adaptation in NMT
  4. Low-Resource NMT
  5. Data
  6. Parallel Source-Domain Data
  7. Target-Domain Data
  8. Dictionaries
  9. Languages
  10. Experimental Setup
  11. mBART Baseline
  12. DALI
  13. LeCA
  14. CPT
  15. Combined
  16. ...and 16 more sections

Figures (2)

  • Figure 1: In our work, which looks at the (previously neglected) intersection of low-Resource NMT and domain adaptation in NMT, we consider only these commonly accessible resources.
  • Figure 2: The trend of averaged sentence-level BLEU (left) and ChrF (right) scores against the token length of the reference translation for DALI models. The scores are averaged across all the model outputs of the same length -- including averaging across languages, where relevant.