Transformer-Based Multilingual NLP Model for Low-Resource Language Translation

Low-resource language machine translation remains an ongoing issue because of the limited parallel corpora and restricted linguistic diversity. While existing multilingual models such as mBERT or XLM-RoBERTa attain high performance on high-resource languages, they do not reliably characterize the morphological, syntactic, or code-switching features of low-resource languages. The contribution of the proposed transformer-based multilingual natural language processing (NLP) model is based on several innovative methodologies and adaptation strategies that include a unique combination of back translation, nCr-IGAN-based data augmentation, and multilingual subword tokenization for effective generalization across diverse linguistic structures. Moreover, contextual embeddings and dimensionality reduction techniques (e.g., GC-DOA and PCA) were adapted for the extraction and optimization of linguistic features. The proposed multilingual model knows how to learn complex syntactic and semantic structures in low-resource language pairs and exhibits low Mean Square Error (MSE) values, moderate cosine similarity, and high BERTScore in the context of English–Hindi translation tasks. High BERTScore indicates a strong level of semantic alignment between the source and target languages, representing an important contribution toward developing comprehensive, high-value multilingual translation systems for endangered language regimes in NLP. When compared with current multilingual models such as mBERT and XLM-RoBERTa, the system shows numerical improvements, particularly regarding BERTScore and cosine similarity, on English–Hindi pairs. These are new, promising, and useful types of architecture that can deal with the complexities of capturing cross-lingual semantics in low-resource environments.