Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dialectal and Low-Resource Machine Translation for Aromanian

Alexandru-Iulius Jerpelea, Alina Rădoi, Sergiu Nisioi

TL;DR

This work tackles machine translation for Aromanian, an endangered language, by assembling a large, multi-genre parallel corpus (79k sentence pairs) and adapting state-of-the-art models. It introduces a language-agnostic Aromanian sentence embedder (LaBSE) and compares NLLB-based MT with several large language models, showing that NLLB variants consistently outperform LLM-based approaches for Aromanian translation. The study also explores orthography normalization (Cunia to DIARO), synthetic data augmentation, and an online translation system (AroTranslate) with CPU-friendly quantization. Results indicate that diversified data and multilingual transfer from Romanian improve Aromanian translation, supporting language preservation while revealing remaining challenges in dialectal variation and data representativeness.

Abstract

This paper presents the process of building a neural machine translation system with support for English, Romanian, and Aromanian - an endangered Eastern Romance language. The primary contribution of this research is twofold: (1) the creation of the most extensive Aromanian-Romanian parallel corpus to date, consisting of 79,000 sentence pairs, and (2) the development and comparative analysis of several machine translation models optimized for Aromanian. To accomplish this, we introduce a suite of auxiliary tools, including a language-agnostic sentence embedding model for text mining and automated evaluation, complemented by a diacritics conversion system for different writing standards. This research brings contributions to both computational linguistics and language preservation efforts by establishing essential resources for a historically under-resourced language. All datasets, trained models, and associated tools are public: https://huggingface.co/aronlp and https://arotranslate.com

Dialectal and Low-Resource Machine Translation for Aromanian

TL;DR

This work tackles machine translation for Aromanian, an endangered language, by assembling a large, multi-genre parallel corpus (79k sentence pairs) and adapting state-of-the-art models. It introduces a language-agnostic Aromanian sentence embedder (LaBSE) and compares NLLB-based MT with several large language models, showing that NLLB variants consistently outperform LLM-based approaches for Aromanian translation. The study also explores orthography normalization (Cunia to DIARO), synthetic data augmentation, and an online translation system (AroTranslate) with CPU-friendly quantization. Results indicate that diversified data and multilingual transfer from Romanian improve Aromanian translation, supporting language preservation while revealing remaining challenges in dialectal variation and data representativeness.

Abstract

This paper presents the process of building a neural machine translation system with support for English, Romanian, and Aromanian - an endangered Eastern Romance language. The primary contribution of this research is twofold: (1) the creation of the most extensive Aromanian-Romanian parallel corpus to date, consisting of 79,000 sentence pairs, and (2) the development and comparative analysis of several machine translation models optimized for Aromanian. To accomplish this, we introduce a suite of auxiliary tools, including a language-agnostic sentence embedding model for text mining and automated evaluation, complemented by a diacritics conversion system for different writing standards. This research brings contributions to both computational linguistics and language preservation efforts by establishing essential resources for a historically under-resourced language. All datasets, trained models, and associated tools are public: https://huggingface.co/aronlp and https://arotranslate.com

Paper Structure

This paper contains 29 sections, 3 figures, 19 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Collecting a Multigenre Dataset
  4. Dataset Split
  5. Synthetic Data
  6. Orthography
  7. Aromanian Sentence Embeddings
  8. Machine Translation Models
  9. GPT Baseline
  10. NLLB Machine Translation
  11. LLMs for Machine Translation
  12. Evaluation
  13. Automatic Evaluation
  14. Human Evaluation
  15. Online Translation System
  16. ...and 14 more sections

Figures (3)

  • Figure 1: Detecting text bounding boxes in "A century of Aromanian poetry".
  • Figure 2: Average loss and accuracy during the LaBSE fine-tuning.
  • Figure 3: AroTranslate GUI.