Context-Aware Neural Machine Translation For English–Hindi Sentences

Abstrak

The increasing demand for accurate and fluent translation between linguistically distant languages, such as English and Hindi, has intensified research in Neural Machine Translation (NMT). However, conventional NMT systems primarily focus on sentence-level translation and often fail to capture broader contextual dependencies essential for preserving meaning, tone, and coherence across sentences. This paper proposes a Context-Aware Neural Machine Translation (CANMT) framework for English–Hindi translation that integrates inter-sentential and intra-sentential context to enhance translation quality. The proposed model employs a Transformer-based architecture augmented with contextual embeddings derived from pre-trained multilingual language models such as mBERT and mBART, enabling effective semantic alignment and contextual representation learning. To handle linguistic divergence, morphological complexity, and word order variation between English and Hindi, the system incorporates hierarchical attention mechanisms and context fusion layers for dynamic context modeling. The model is trained and fine-tuned on a large-scale English–Hindi parallel corpus, followed by rigorous evaluation using BLEU, METEOR, and TER metrics. Experimental results demonstrate that the CANMT model achieves superior performance compared to traditional sequence-to-sequence and vanilla Transformer models, yielding improvements in contextual coherence, fluency, and semantic fidelity. The findings underscore the importance of incorporating contextual information in neural architectures for enhancing low-resource and morphologically rich language translation tasks, paving the way for more intelligent and human-like multilingual communication systems.

Identifier

• OpenAlex: https://openalex.org/W7163257756
• DOI: https://doi.org/10.25258/ijddt.16.43s.113