過往研究中，情感分析在藥物推薦系統中的應用多集中於將使用者評論轉換為單一情感分數，然而此種單一表示方式往往無法完整捕捉評論中蘊含的多元情感面向，進而影響推薦結果的準確性。此外，傳統推薦系統多以黑盒方式運作，僅提供推薦結果，卻無法說明推薦原因。在藥物這類高度敏感與決策風險較高的領域，缺乏可解釋性的推薦方式，可能導致病患對系統的信任度降低。 本研究提出一套結合主題建模與情感分析的架構，從藥物的評論中提取多維度情感資訊，並以此提升基於內容的推薦品質。首先，考量在含有評論與評分的資料集中，常見評分與評論內容不一致的現象，研究首先探討所提出之異常值處理策略的合理性與最佳組態，並選出最適用的情感分析方法。 接著，本研究透過LDA（Latent Dirichlet Allocation），從多個角度確定各資料集的最佳主題數，進而建構每個藥物的多維度情感陣列。隨後，將此導入基於內容的推薦演算法，並嘗試三種偏好向量建構方式（Average、Weighted Average、Max Pooling），分別評估其在短列表（N = 2~5）與長列表（N = 10、15）下的NDCG＠N表現。實驗結果顯示，Max Pooling在短列表推薦中表現最佳，但在長列表中效能明顯下滑；而Weighted Average雖在短列表中次於Max Pooling，卻能在長列表下持續維持穩定表現，為整體最具實用性的向量建構方法，而多維度情感陣列的推薦品質皆可維持在七成以上，優於傳統單一情感分數方法。 最後，本研究亦探討大型語言模型（Large Language Models, LLM）在推薦理由生成上的應用，透過實驗中產出的數據，結合LLM生成具主題性與摘要性的推薦理由，進一步提升推薦系統的可解釋性與使用者理解度。 ;In earlier studies, drug recommendation systems often used just one overall sentiment score from user reviews. However, this simple approach misses the detailed feelings in the reviews and can reduce recommendation accuracy. Also, traditional systems often lack clear explanations, which may lower user trust, especially in medical settings. To improve this, our study combines topic modeling and sentiment analysis to get a more complete view of emotions in drug reviews. We first tested a method to remove unusual data caused by mismatches between ratings and review content and then picked the best sentiment analysis method. Next, we used LDA (Latent Dirichlet Allocation) to find the best number of topics for each dataset and created a multi-dimensional sentiment profile for each drug. This profile was used in a content-based recommendation system, and we tested three ways to build preference vectors: Average, Weighted Average, and Max Pooling. Results showed Max Pooling worked best for short lists (2–5 items), while Weighted Average was more stable and better for longer lists (10–15 items). Finally, we used a Large Language Model (LLM) to generate clear recommendation reasons. By using topic and sentiment data, the LLM was able to produce easy-to-understand summaries, helping users trust and understand the system better.
