The Research on Military‐Retrieval‐Augmented Generation Question Answering System Based on MiliGNER

With the rapid development of large language models, generative named entity recognition and retrieval‐augmented generation have opened new methodological pathways for intelligent question answering in specialised domains. To address the characteristics of military texts, including complex entity naming, tightly coupled attribute dependencies and highly implicit semantic expressions, this study proposes a military‐RAG framework for military question–answering tasks and designs a generative named entity recognition module, termed MiliGNER, as the core component for structured extraction. The proposed method formulates entity recognition as a sequence generation process from input text to structured output, directly generating entity–attribute–value triplets to enhance the representation of complex military entities and their attribute relationships. To improve the quality of structured extraction and downstream question answering, the proposed framework is optimised from three perspectives: structural consistency, semantic consistency and question–answering correctness. Specifically, joint training with positive and negative samples enhances the model's ability to discriminate invalid inputs and nontriggering cases; an attribute validation module is introduced to reduce attribute mismatches and semantic drift; and entity‐aware attention structure‐guided decoding, together with a multipath error‐correction mechanism, improve the completeness and robustness of structured generation in complex scenarios. On this basis, military‐RAG uses validated structured triplets for retrieval prompt construction and answer generation, thereby improving the answer reliability of retrieval inputs and the semantic alignment between the final answer and the original question. Experiments were conducted on both public benchmark datasets and a self‐constructed military‐domain dataset. The results show that, under the current experimental setting, MiliGNER achieves consistent improvements over the selected comparison methods in structured entity recognition tasks, whereas military‐RAG also demonstrates strong overall performance in question–answering tasks relative to representative reference systems. Ablation studies further indicate that each key module makes a practical contribution to overall system performance. Overall, the proposed method provides an effective collaborative framework of structured modelling and generative reasoning for highly specialised question–answering tasks in the military domain.