Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Pivot Language for Low-Resource Machine Translation

Abhimanyu Talwar, Julien Laasri

TL;DR

This study tackles low-resource Nepali-English translation by introducing Hindi as a pivot language. It evaluates a fully supervised Transfer method and a semi-supervised Backtranslation approach, demonstrating a substantial improvement over a prior fully supervised baseline with a devtest SACREBLEU of $14.2$ for the transfer pipeline. The paper compiles diverse Nepali–Hindi and Hindi–English corpora, leverages large Hindi monolingual data, and analyzes pivot-related phenomena such as historical relatedness, reordering, and morphology. Overall, pivot-based transfer yields strong gains, though backtranslation can introduce noise, motivating future work on synthetic data strategies and extended training to close the gap with semi-supervised baselines and further improve robustness.

Abstract

Certain pairs of languages suffer from lack of a parallel corpus which is large in size and diverse in domain. One of the ways this is overcome is via use of a pivot language. In this paper we use Hindi as a pivot language to translate Nepali into English. We describe what makes Hindi a good candidate for the pivot. We discuss ways in which a pivot language can be used, and use two such approaches - the Transfer Method (fully supervised) and Backtranslation (semi-supervised) - to translate Nepali into English. Using the former, we are able to achieve a devtest Set SacreBLEU score of 14.2, which improves the baseline fully supervised score reported by (Guzman et al., 2019) by 6.6 points. While we are slightly below the semi-supervised baseline score of 15.1, we discuss what may have caused this under-performance, and suggest scope for future work.

Pivot Language for Low-Resource Machine Translation

TL;DR

This study tackles low-resource Nepali-English translation by introducing Hindi as a pivot language. It evaluates a fully supervised Transfer method and a semi-supervised Backtranslation approach, demonstrating a substantial improvement over a prior fully supervised baseline with a devtest SACREBLEU of for the transfer pipeline. The paper compiles diverse Nepali–Hindi and Hindi–English corpora, leverages large Hindi monolingual data, and analyzes pivot-related phenomena such as historical relatedness, reordering, and morphology. Overall, pivot-based transfer yields strong gains, though backtranslation can introduce noise, motivating future work on synthetic data strategies and extended training to close the gap with semi-supervised baselines and further improve robustness.

Abstract

Certain pairs of languages suffer from lack of a parallel corpus which is large in size and diverse in domain. One of the ways this is overcome is via use of a pivot language. In this paper we use Hindi as a pivot language to translate Nepali into English. We describe what makes Hindi a good candidate for the pivot. We discuss ways in which a pivot language can be used, and use two such approaches - the Transfer Method (fully supervised) and Backtranslation (semi-supervised) - to translate Nepali into English. Using the former, we are able to achieve a devtest Set SacreBLEU score of 14.2, which improves the baseline fully supervised score reported by (Guzman et al., 2019) by 6.6 points. While we are slightly below the semi-supervised baseline score of 15.1, we discuss what may have caused this under-performance, and suggest scope for future work.

Paper Structure

This paper contains 23 sections, 2 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Nepali-Hindi Parallel Corpus
  4. Hindi-English Parallel Corpus
  5. Monolingual data
  6. Related Work
  7. Transfer method
  8. Synthetic method
  9. Triangulation method
  10. Models
  11. Fully-Supervised
  12. Semi-Supervised
  13. Hindi as a Pivot Language
  14. Historical Relatedness
  15. Amount of Reordering
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Examples of ill-formed sentences from the Nepali-English Parallel Corpus.
  • Figure 2: The amount of word reordering required is minimal when translating between Hindi and Nepali but is significant when translating between Nepali and English.
  • Figure 3: Nepali and Hindi share several morphological features and are different from English.