GraphSageNMT: Graph-Augmented Neural Machine Translation With Dynamic Context Retrieval for Indic Languages

Translating ancient Indic texts into English poses significant challenges for neural machine translation (NMT) systems due to their brevity, distinct grammatical structures, and inherent complexity in interpretation. The current work proposes GraphSageNMT, a memory-augmented NMT framework that combines multi-level linguistic features with a non-parametric, memory-augmented module that can be used at any time. GraphSageNMT uses a graph-based semantic memory and wisdom/inference that is derived from associative learning. Pre-trained language models are used to combine contextual embeddings with explicit linguistic annotations, such as Part-of-Speech tags, dependency linkages, and semantic role labels, via a residual gating-based multi-feature fusion mechanism. In addition to improving local features, a two-part external memory system is added. The first part is a graph-based semantic memory layer that builds document-level associative relationships between concepts. The second part is a lightweight long-contextual memory buffer for non-parametric continual learning. Cross-attention in the decoder combines memory-retrieved contexts with current representations. This allows the model to maintain both syntactic fidelity and semantic continuity across passages. A position-aware upgrade, with dropout regularisation, is applied to make our model more robust and better at generalising. Empirical research on the proposed GraphSageNMT approach shows significant improvements in Indic-English language performance across various fully automatic evaluation metrics. The results show that dynamic memory retrieval and multi-feature conditioning are effective at capturing the subtle and associative aspects of translating texts. This is a step toward NMT systems that are more human-like and can continuously learn.