HSGraphAgent: Knowledge-Graph-Guided Large Language Models for Harmonized System Code Classification
HSGraphAgent knowledge-graph-guided reasoning framework. Source: paper.摘要
HSGraphAgent represents the Harmonized System hierarchy and exclusion notes as an explicit knowledge graph that guides large language models through legal classification paths. Its Select-Redirect mechanism restricts each decision to valid child nodes and redirects reasoning when an exclusion rule is triggered. Experiments on four-digit headings and six-digit HS codes show that rule-aware graph reasoning substantially improves fine-grained accuracy and interpretability over direct generation and retrieval-augmented baselines.
出版物
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, 44761-44773
研究概览
HSGraphAgent 将海关 HS 编码体系建模为包含层级关系和排除规则的知识图谱，引导大语言模型从 Section、Chapter、Heading 到 Subheading 逐级完成商品归类。该方法强调规则约束与可解释推理，避免仅凭文本相似度选择编码。
核心方法与发现
- Select 阶段只在当前节点的合法子节点中选择，保证层级路径有效。
- Redirect 阶段在触发税则排除条件时动态修正推理路径。
- 在 6 位 HS Code 细粒度任务上，图谱引导的分层推理明显优于直接生成和常规 RAG 基线。
资料来源
夏强
博士研究生
王奥
硕士毕业生
