本研究旨在開發一套結合大型語言模型（Large Language Model, LLM）與知識圖譜（Knowledge Graph, KG）的投資機器人系統，以解決卷積神經網路（Convolutional Neural Network, CNN）與LLM預測模型在決策解釋性上的不足。系統先採用LLM對券商研究報告進行摘要與分析漲跌方向作為定性特徵，並以技術分析工具TA-Lib（Technical Analysis Library）計算八項常見技術指標作為定量特徵。兩者融合後輸入CNN模型進行股票漲跌方向之預測，並將預測結果及輸入特徵轉為三元組（Triple）形式儲存至KG，為支援自然語言查詢預測依據與技術指標資料等，系統結合LangChain框架與圖檢索增強生成技術（Graph Retrieval-Augmented Generation, GraphRAG），提升查詢透明度與互動性。 設計上，本研究運用MIAT（Modular Integrated Architecture Theory）方法論，將系統分為多個功能模組，強化架構靈活性與實作效率。實驗結果顯示，解決LLM預測模型的幻覺現象（Hallucination），並證實融合定性與定量特徵後的CNN模型的分類Accuracy可由0.43提升至0.54、Macro F1 Score與Macro ROC-AUC（Macro-averaged Receiver Operating Characteristic Area Under Curve）均明顯優於僅使用定量特徵的模型，且自然語言查詢平均約9秒即可回答問題。 綜合而言，本研究建構之系統不僅降低操作門檻，亦具備可追溯與解釋決策之能力，為LLM在投資輔助應用提供參考模型。 ;This study develops an investment robot system integrating Large Language Models (LLMs) and Knowledge Graphs (KGs) to enhance the interpretability of predictions from Convolutional Neural Networks (CNNs) and LLM-based models in decision-making. The system employs an LLM to summarize brokerage research reports and extract upward/downward movement tendencies as qualitative features, while using the Technical Analysis Library (TA-Lib) to compute eight common technical indicators as quantitative features. These features are fused and input into a CNN model for stock movement direction classification, with the input features and prediction results represented as triples and stored in the KG. To enable natural language queries for prediction rationale and technical indicator data, the system integrates the LangChain framework with Graph Retrieval-Augmented Generation (GraphRAG). This approach enhances query transparency and user interactivity. Designed using the Modular Integrated Architecture Theory (MIAT), the system is divided into modular components for improved flexibility and implementation efficiency. Experimental results show that the system mitigates hallucination in LLM-based models. The CNN model′s classification accuracy improves from 0.43 to 0.54 with the addition of qualitative features, and both the Macro F1 score and the Macro ROC-AUC (Macro-averaged Receiver Operating Characteristic Area Under Curve) are significantly better than those of the model using only quantitative features. Natural language queries are answered in approximately 9 seconds on average. Overall, the system lowers the learning barrier for non-expert users, enables traceable and explainable decision support, and serves as a reference model for applying LLMs and KGs in investment decision-support systems.
