Read the full-text of this article for free

Title: LDCIR-Trans: a lightweight dependency-constrained iterative refinement model for machine translation

Authors: Mengyao Wang; Xihe Qiu

Addresses: School of Foreign Languages, Shanghai University of Engineering Science, Shanghai, 201620, China ' School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China

Abstract: While neural machine translation facilitates effective cross-lingual information transfer, existing lightweight architectures continue to encounter significant challenges in structural modelling precision and decoding stability. They struggle with long-range and syntactic dependencies, shallow attention limits fine-grained structure representation, and compressed architectures often cause semantic drift or repetition. To address these issues, we propose LDCIR-Trans, a lightweight structure-aware translation model. It introduces essential structural priors and a stable decoding mechanism while remaining compact. First, a dependency graph modelling (DGM) module explicitly constructs dependency graphs to supply syntactic cues and compensate for limited global modelling capacity. Second, a dependency-constrained iterative refinement (DCIR) mechanism guides decoding with structure-enhanced signals, enables progressive correction, and reduces semantic deviation. Finally, the lightweight structure-aware decoder (LSAD) employs parameter sharing and distribution calibration to improve representation and stability. Experiments show that LDCIR-Trans achieves high efficiency and significantly outperforms existing lightweight baselines.

Keywords: neural machine translation; NMT; dependency graph modelling; DGM; gate-augmented iterative update; lightweight structure-aware decoder; LSAD.

DOI: 10.1504/IJICT.2026.153518

International Journal of Information and Communication Technology, 2026 Vol.27 No.46, pp.80 - 102

Received: 10 Dec 2025
Accepted: 08 Feb 2026
Published online: 12 May 2026 *