Segmentation-Aware and Context-Reranked Retrieval-Augmented Generation for Vietnamese Question Answering

Retrieval-augmented generation (RAG) enhances knowledge-intensive question answering by integrating document retrieval with neural generation. However, retrieval performance in Vietnamese and other low-resource languages is often limited by word segmentation ambiguity and semantic mismatch. This paper proposes SADR-CoRe-Gen, a three-stage RAG framework comprising (i) segmentation-aware dense retrieval (SADR), which combines subword- and phrase-level representations for robust retrieval; (ii) context-reranked retrieval (CoRe), which refines retrieved passages through query-passage interactions; and (iii) a fine-tuned ViT5 generator for answer generation. The framework is evaluated on the ViQuAD2 and ViHealthQA benchmarks using 10 independent runs with different random seeds. Experimental results show that SADR improves first-stage retrieval robustness, while CoRe provides substantial additional gains through contextual reranking. Their integration in SADR-CoRe-Gen achieves highly competitive performance, with best or tied-best results on several evaluation metrics across the two benchmark datasets. On ViHealthQA, it achieves [Formula: see text], [Formula: see text], and ROUGE-L [Formula: see text] [Formula: see text]. On ViQuAD2, it achieves the highest [Formula: see text] and ROUGE-L [Formula: see text] [Formula: see text]. Statistical analyses using the Friedman test and the Wilcoxon signed-rank test confirm significant improvements over weaker retrieval baselines, demonstrating the effectiveness and robustness of combining segmentation-aware retrieval with contextual reranking for Vietnamese RAG.