面向區塊鏈漏洞知識庫的大模型增強知識圖譜問答模型

摘" 要： 大語言模型（LLM）在專業領域特別是區塊鏈漏洞領域應用時存在局限性，如專業術語噪聲干擾和細粒度信息過重導致理解不足。為此，構建一種面向區塊鏈漏洞知識庫的增強型知識圖譜問答模型（LMBK_KG）。通過整合大模型和知識圖譜來增強知識表示和理解能力，同時利用多粒度語義信息進行專業問題的過濾和精準匹配。研究方法包括使用集成的多粒度語義信息和知識圖譜來過濾專業術語噪聲，以及采用大模型生成的回答與專業知識圖譜進行結構化匹配和驗證，以提高模型的魯棒性和安全性。實驗結果表明，所提出的模型在區塊鏈漏洞領域問答的準確率比單獨使用大模型提高26%。

關鍵詞： 大語言模型； 知識圖譜； 問答模型； 多粒度語義信息； 區塊鏈； 漏洞信息； 文本表征

中圖分類號： TN929.5?34； TP391.1" " " " " " " " 文獻標識碼： A" " " " " " " " " " 文章編號： 1004?373X（2025）02?0137?06

Large model enhanced knowledge graph question answering model for blockchain vulnerability knowledge base

XIE Fei1， SONG Jianhua2， 5， JIANG Li1， ZHANG Yan1， 4， HE Shuai3

（1. School of Computer Science and Information Engineering， Hubei University， Wuhan 430062， China;

2. School of Cyber Science and Technology， Hubei University， Wuhan 430062， China;

3. School of Cyber Science and Engineering， Huazhong University of Science and Technology， Wuhan 430074， China;

4. Key Laboratory of Intelligent Sensing System and Security （Hubei University）， Ministry of Education， Wuhan 430062， China;

5. Hubei Provincial Engineering Research Center of Intelligent Connected Vehicle Network Security， Wuhan 430062， China）

Abstract： There are limitations in the application of large language models （LLMs） in professional fields， especially in the field of blockchain vulnerabilities， such as noise interference of technical terms and insufficient understanding caused by excessive fine?grained information. On this basis， an enhanced knowledge graph question answering model for blockchain vulnerability knowledge base （LMBK_KG） is constructed， which can enhance the knowledge representation and comprehension ability by integrating large models and knowledge graphs， and filter and accurately match professional problems by means of multi?granularity semantic information. The research methods include using integrated multi?granularity semantic information and knowledge graph to filter the professional term noise， and using large model?generated answers for structured matching and validation with the professional knowledge graph to improve the robustness and security of the model. The experimental results show that， in comparison with the large model used alone， the proposed model can improve the accuracy of question answering in the field of blockchain vulnerabilities by 26%.

Keywords： large language model; knowledge graph; question?answering model; multi?granularity semantic information; blockchain; vulnerability information; text representation

0" 引" 言

隨著大語言模型（LLM）[1]時代的到來，自然語言處理（NLP）領域經歷了革命性的變革，其中LLM如InstructGPT[2]、ChatGPT和GPT?4[3]等在廣泛的問答任務中展現出了卓越的性能。這些模型能夠理解并執行復雜的人類指令，準確解答各類問題。盡管如此，它們在處理特定領域，尤其是充滿專業術語的區塊鏈漏洞領域時，效果并不總是理想。針對上述問題，本文提出一種創新的解決方案，即面向區塊鏈漏洞知識庫的大模型增強知識圖譜問答模型（LMBK_KG）。該模型結合了大型語言模型和多層次語義信息，利用區塊鏈專業知識庫來提升對區塊鏈漏洞問題的回答精度[4]。

1" 模型概述