AgE-Rerank:面向农业病害问答的实体感知 RAG 文档重排序方法

    AgE-Rerank: An entity-aware RAG document reranking method for agricultural disease question answering

    • 摘要:
      目的 在农业病害智能问答场景中,防治方案高度依赖作物、病害、生育期和药剂等实体信息,对答案的证据可靠性与情境适配性要求较高。检索增强生成(Retrieval-augmented generation, RAG)通过引入外部植保知识库增强生成内容的可解释性,但普遍存在检索文档主题相关而关键实体失配的问题,易产生不可靠的防治推荐。为此,本文提出实体感知文档重排序方法AgE-Rerank(Agricultural entity-aware reranker)。
      方法 该方法以 jina-reranker-v3 为基础,结合农业命名实体识别、实体注意力对齐和候选集实体重加权,融合全局语义表示与实体交互表示计算相关性分数,将关键农业实体与问题相符的证据排至前列。
      结果 在自建农业病害 RAG 数据集上,AgE-Rerank 的 Recall@1、MRR@10 和 RAG-F1@3 分别达到 0.7560、0.8362 和 0.9274,优于 jina-reranker-v3 等基线;消融和案例分析表明,实体注意力对齐与候选集实体重加权均能减少关键实体不匹配造成的证据错排,并改善生成答案质量。
      结论 AgE-Rerank 通过显式建模问题与文档间的农业实体匹配关系,提高了证据排序的准确性,为农业病害智能问答提供了方法支撑。

       

      Abstract:
      Objective In the scenario of intelligent question answering for agricultural diseases, disease control strategies are highly dependent on entity information such as crops, diseases, growth stages, and pesticides. This imposes strict requirements on the reliability of evidence and the contextual adaptability of responses. Retrieval-augmented generation (RAG) enhances the interpretability of generated content by incorporating external plant protection knowledge bases. However, it commonly suffers from key entity mismatches in topically relevant retrieved documents, which tends to yield unreliable control recommendations. To address this issue, an agricultural entity-aware reranking method named AgE-Rerank (Agricultural entity-aware reranker) was proposed.
      Method Built on jina-reranker-v3, the method integrates agricultural named entity recognition, entity attention alignment, and candidate-set entity reweighting. It fuses global semantic and entity-interaction representations to calculate relevance scores and prioritizes evidence whose key agricultural entities match the query.
      Result On a self-constructed agricultural disease RAG dataset, AgE-Rerank achieved Recall@1, MRR@10, and RAG-F1@3 scores of 0.7560, 0.8362, and 0.9274, respectively, outperforming baselines including jina-reranker-v3. Ablation and case analyses showed that both entity attention alignment and candidate-set entity reweighting reduced evidence misranking caused by mismatched key entities and improved the quality of generated answers.
      Conclusion By explicitly modeling agricultural entity matching between questions and documents, AgE-Rerank improves evidence-ranking accuracy and provides methodological support for intelligent question answering about agricultural diseases.

       

    /

    返回文章
    返回