| 摘要: | 傳統的推薦系統主要依賴於分析數據和機器學習算法，並由系統單方面向使用者推播。對話式推薦系統則可以直接接受來自使用者主動提供的資訊，而系統也能透過文字對項目進行推薦，給予最直接的幫助。
在對話式推薦系統（Conversational Recommender System, CRS）中使用大型語言模型（Large Language Model, LLM）能獲得許多傳統模型無法擁有的優勢。首先，LLM的系統不需要經過訓練即可展現出色的性能，能夠解決冷啟動問題。其次，LLM的泛用性及可擴展性極高，能適應或導入到各種應用場景。
大多數以往的研究偏向於從大量的物件中推薦出一個最相關的內容，這較適合使用檢索增強生成（Retrieval Augmented Generation, RAG）技術；我們的研究著重於從少量的物件中推薦一個最適合的內容，並強調應依據病因來推薦藥方。
本研究首次嘗試以失眠患者的中藥方推薦作為任務目標，使用繁體中文能力出色的gemini-1.5-flash模擬並生成病患與中醫師之間的對話。我們提出Hint Module來導入生理量測及中醫問診技巧，透過偵測特定字串來觸發系統機制，給予LLM額外的提示訊息藉此影響它的輸出結果。
我們的實驗結果顯示，此方法可生成寫實的自述，能被視作為良好的對話範例，並在十種中藥方的推薦任務中，可得到八成以上的準確率及Macro-F1成績。其中Hint Module能顯著地改善多輪對話後的成績表現 (p-value < 0.01)。
最後我們也以中醫學的各項觀點進行分析，透過視覺化的圖表呈現出各個藥方之間分布上的關聯性，以得到更全面及清晰的瞭解。
實驗結果展示了以LLM打造的中藥方推薦系統擁有傑出的基礎能力，並擁有絕佳的可擴展性。
;Traditional recommender systems primarily rely on data analysis and machine learning algorithms, with recommendations being pushed to users unilaterally by the system. In contrast, conversational recommender systems can directly accept information actively provided by users, and the system can recommend items through text, offering the most direct assistance.
Using a Large Language Model (LLM) in a Conversational Recommender System (CRS) offers many advantages that traditional models do not have. First, an LLM system does not require training to perform well, which solves the cold-start problem. Second, LLMs possess high versatility and scalability, allowing them to adapt or be integrated into various application scenarios.
Most previous studies focused on recommending the most relevant content from a large set of items, which is more suitable for Retrieval Augmented Generation (RAG) techniques. Our research, however, focuses on recommending the most appropriate content from a small set of items, emphasizing the recommendation of prescriptions based on the cause of illness.
Our study is the first to attempt using Traditional Chinese Medicine (TCM) prescription recommendations for insomnia patients as a task goal, utilizing gemini-1.5-flash, which excels in Traditional Chinese, to simulate and generate conversations between patients and TCM doctors. We propose using a Hint Module to incorporate physiological measurements and consultation techniques, triggering the system to provide LLM with additional prompt messages by detecting specific strings, thereby influencing the LLM’s output.
Our experimental results show that this method can generate realistic patient statements that can be regarded as good examples, achieving over 80% accuracy and Macro-F1 score in recommending ten classes of prescriptions. The Hint Module can significantly (p-value < 0.01) improve the performance of the multiple rounds of dialogue.
Lastly, we conducted an analysis based on various perspectives of TCM, we used visualized charts to present the distributional relationships among different prescriptions, aiming for a more comprehensive and clearer understanding.
The experimental results show that the TCM prescription recommender system built with LLM has excellent basic capabilities and scalability. |
