A Semantic-Enhanced Multi-Task Framework with Structure-Aware Curriculum Learning for Low-Resource Neural Machine Translation

Abstract Adapting large language models to low-resource neural machine translation remains challenging due to semantic misalignment and optimization interference during multi-task fine-tuning. To address these issues, this study proposes a semantic-enhanced generative multi-task learning framework with dynamic curriculum scheduling, built on Qwen2-7B-Instruct and trained via parameter-efficient LoRA adaptation. We reformulate Semantic Role Labeling (SRL) as a generative auxiliary task which, together with named entity recognition and related-language transfer, enables a decoder-only LLM to encode predicate-argument structures within a shared adapter space and thereby improve structural consistency. We further introduce a structural difficulty metric that combines predicate, argument, and argument-span complexity with initial model loss, and use it to drive curriculum learning, progressively expanding the sampling pool from syntactically simple to structurally complex instances and stabilizing early-stage optimization. Experimental results demonstrate consistent improvements when integrating related-language transfer and generative SRL. On Chinese-to-Lao (Zh-Lo) translation, our method achieves a +5.58 BLEU improvement over a strong supervised baseline and outperforms massively multilingual and region-adapted baselines under matched data. Training-dynamics analysis shows that the semantic-driven curriculum alleviates early gradient conflicts and training plateaus. Cross-lingual evaluation on Chinese-to-Myanmar (Zh-My) yields a further +1.77 BLEU gain, confirming the robustness and transferability of the framework across typologically similar isolating languages.