GAN-based Incremental Learning for English to Khasi Machine Translation: A Low-Resource Language Model

Abstract

Design of a highquality machine translation system for lowresource languages is still a big problem because there are not sufficient parallel corpora and the patterns of languages frequently shift In order to address such a gap this paper presents a modified generative adversarial networkbased hybrid translation framework known as the Incremental Generative Ad versarial Network Integrated Hybrid approach IGANH It provides incremental learning and robust translation for the EnglishKhasi language pair This framework consists of an incremental bilingual data processor that updates vocabulary and phrase alignments in real time A hybrid translation core that integrates statistical alignment priors with a transformerbased encoder–decoder to improve syntactic and semantic grounding A modified conditional GAN that generates highquality pseudoparallel sentences and learns new linguistic patterns over time and an adaptive optimization controller that balances adversarial learning signals and avoids catastrophic errors during model updates It can update the newly generated representations and GAN components when new data is generated This makes it more efficient and maintains performance consistency between training iterations Experimental evaluations indicate that the suggested framework improves trans lation consistency increases the coverage of rare language structures and provides a scalable approach for constructing translation systems in lowresource settings Overall the proposed framework offers a flexible option to implement machine translation that remains effective in changing language settings