Abstract: To address the challenges of insufficient label semantics understanding, vague relationship modeling, and high computational costs in zero-shot re-ranking tasks for large language models (LLMs), a hierarchical filtering and label semantics extension approach named HFLS was proposed. A multi-level label semantic extension path was constructed, and a progressive prompting strategy ("keyword matching → semantic association → domain knowledge integration") was designed to guide LLM in deep relevance reasoning. A hierarchical filtering mechanism was introduced to reduce computational complexity while retaining high-potential candidate documents. Experimental results indicate that on seven benchmark datasets (e.g., TREC DL 2019), HFLS achieves average gains of 21.92%, 13.43% and 8.59%in NDCG@10 compared to Pointwise methods like Pointwise.qg, Pointwise.yes_no, and Pointwise.3Label. In terms of inference efficiency, the processing latency per query is reduced by 91.06% compared to Listwise methods, 68.87% compared to Pairwise methods, and 33.54% compared to Setwise methods.

Key words: large language model, zero-shot learning, re-ranking, information retrieval, prompt engineering.

摘要： 针对大语言模型(LLM)在零样本重排序任务中存在的标签语义理解不足、关系建模模糊及计算成本过高问题，提出基于层次过滤与标签语义扩展的HFLS重排序方法。通过构建多级标签语义扩展路径，设计“关键词匹配→语义关联→领域知识整合”的递进式提示策略，引导LLM实现深度相关性推理；同时引入分层过滤机制，在降低计算复杂度的同时保留高潜力候选文档。实验结果表明：在TREC DL 2019等7个基准数据集上，HFLS相较于Pointwise.qg、Pointwise.yes_no和Pointwise.3Label等Pointwise方法，NDCG@10指标平均提升21.92%、13.43%、8.59%。推理效率方面，单个查询处理时延较Listwise方法降低91.06%，较Pairwise方法降低68.87%，较Setwise方法降低33.54%。

关键词: 大语言模型, 零样本学习, 重排序, 信息检索, 提示工程