Open Access

ARTICLE

PNMT: Zero-Resource Machine Translation with Pivot-Based Feature Converter

Lingfang Li^1,2, Weijian Hu², Mingxing Luo^1,*

1 School of Information Science and Technology, Southwest Jiaotong University, Chengdu, 611730, China
2 School of Information Engineer, Inner Mongolia University of Science & Technology, Baotou, 014000, China

* Corresponding Author: Mingxing Luo. Email:

Computers, Materials & Continua 2025, 84(3), 5915-5935. https://doi.org/10.32604/cmc.2025.064349

Received 12 February 2025; Accepted 27 May 2025; Issue published 30 July 2025

Abstract

Neural machine translation (NMT) has been widely applied to high-resource language pairs, but its dependence on large-scale data results in poor performance in low-resource scenarios. In this paper, we propose a transfer-learning-based approach called shared space transfer for zero-resource NMT. Our method leverages a pivot pre-trained language model (PLM) to create a shared representation space, which is used in both auxiliary source→pivot (Ms2p) and pivot→target (Mp2t) translation models. Specifically, we exploit pivot PLM to initialize the Ms2p decoder and Mp2t encoder, while adopting a freezing strategy during the training process. We further propose a feature converter to mitigate representation space deviations by converting the features from the source encoder into the shared representation space. The converter is trained using the synthetic source→target parallel corpus. The final Ms2t model combines the Ms2p encoder, feature converter, and Mp2t decoder. We conduct simulation experiments using English as the pivot language for German→French, German→Czech, and Turkish→Hindi translations. We finally test our method on a real zero-resource language pair, Mongolian→Vietnamese with Chinese as the pivot language. Experiment results show that our method achieves high translation quality, with better Translation Error Rate (TER) and BLEU scores compared with other pivot-based methods. The step-wise pre-training with our feature converter outperforms baseline models in terms of COMET scores.

---

Keywords

---