基于大语言模型与知识图谱的计算机考研个性化智能实训系统

Personalized Intelligent Training System for Computer Science Postgraduate Entrance Examinations Based on Large Language Models and Knowledge Graphs

摘要: 针对计算机专业考研备考中知识资源碎片化、练习反馈分散以及通用大语言模型领域回答缺少依据的问题，设计并实现了一种融合大语言模型与知识图谱的个性化智能实训系统。系统采用Spring Boot与Vue3构建前后端分离架构，使用MySQL、Redis、Neo4j和对象存储分别存储业务数据、会话缓存、图谱与向量数据以及试卷文件。面向智能问答，提出意图识别驱动的混合检索流程：非专业问题直接进入模型生成链路，专业问题并行执行向量语义召回和知识图谱子图扩展，经重排序后将文本证据、关联实体及来源信息注入提示词，再由Qwen模型生成可追溯回答。系统进一步将试卷数字化、个性化出题、答题记录、错题再练和划线笔记连接为学习闭环。黑盒测试覆盖登录、题目作答、历史记录、智能问答、社区、文件上传和越权访问等核心场景，验证了系统业务链路与权限隔离的可用性。该系统能够统一组织教材、试题、知识点与学习记录，为考生提供具有知识关联、来源展示和持续反馈能力的一站式智能学习环境。

Abstract: To address fragmented knowledge resources, disconnected learning feedback, and the weak evidential grounding of general-purpose large language models in computer science postgraduate entrance examination preparation, this paper designs and implements a personalized intelligent training system that integrates a large language model with a knowledge graph. The system adopts a Spring Boot and Vue 3 architecture and uses MySQL, Redis, Neo4j, and object storage for business data, session caching, graph and vector data, and examination files. For question answering, an intent-driven hybrid retrieval workflow is implemented. General questions are routed directly to the language model, whereas domain questions trigger parallel vector retrieval and knowledge-graph subgraph expansion. After reranking, textual evidence, related entities, and source metadata are injected into the prompt for Qwen to produce traceable answers. The system also connects examination-paper digitization, personalized question generation, answering records, error correction, and note taking into a learning loop. Black-box tests of core scenarios produced outputs consistent with expectations, demonstrating the functional availability of the main workflows and access-control mechanism. The system provides an integrated learning environment with structured knowledge association, source presentation, personalized practice, and continuous learning feedback.