Multilingual neural machine translation (MNMT) plays a crucial role in extending language technologies to underrepresented and linguistically diverse regions. However, existing MNMT systems remain opaque, particularly when applied to structurally divergent and low-resource languages such as those found in Northeast India. This paper proposes an explainability-aware MNMT framework, AGER-MNMT, optimized for English \(\rightarrow \) Indic and Indic \(\rightarrow \) English translation across six low-resource languages: Assamese, Bodo, Khasi, Manipuri, Mizo, and Nepali, spanning the Indo-Aryan, Tibeto-Burman, and Austroasiatic families. The architecture extends a shared transformer backbone with language-conditioned adapter modules, sparsely activated mixture-of-experts (MoE) layers, and a novel Attribution-Guided Expert Router (AGER) that integrates token-level attribution signals into expert routing. This design enables effective parameter sharing while preserving linguistically meaningful specialization across language families. Trained on curated parallel data augmented with back-translation, the model achieves competitive performance, including BLEU scores of 29.8 for Assamese and 26.7 for Nepali under the AGER-enhanced configuration. To improve transparency, we incorporate complementary post-hoc interpretability techniques—attention visualization, SHAP, and LIME—which provide token-level and layer-wise insights into translation behaviour. Quantitative and qualitative analyses show that AGER stabilizes expert utilization, improves representation quality for low-resource languages, and enhances the interpretability of translation decisions. Overall, the results demonstrate that combining adapters, MoE-based specialization, and attribution-guided routing offers a promising path toward both high-performance and explainable MNMT for low-resource Indic languages.
