隨著資訊爆炸挑戰日益嚴峻，人們對從多篇文檔中迅速獲取精簡資訊的需求日益增加，這在新聞總結、學術文獻和電影評論等領域得到了廣泛應用。因此，本研究專注於研究醫療問題的多答案摘要，旨在幫助民眾從大量相關的醫療答案中提煉出可行且可靠的回覆，以便他們能夠更容易地理解。我們提出了一種基於階層式聚類注意力機制 (Hierarchical Clustering with Attentions, HCA) 的多文檔摘要模型。該模型利用修辭結構理論提取每篇文檔的基本話語單元，並進行基於密度的階層式聚類，將這些基本話語單元分成多個群集。隨後，利用大型語言模型找出每個群集的主題，以增強模型對群集之間基本話語單元的理解。通過對群集進行重排序，模型能更好地捕捉聚類後群集間的差異性。最終，我們將不同群集的主題與基本話語單元整合成輸入嵌入向量，並輸入到階層式的編碼器-解碼器架構中，包含階層式自我注意力機制以及階層式交叉注意力機制，生成最終的醫療問題多答案摘要。 有鑒於在多文檔摘要領域缺乏公開的中文醫療領域評測資料集，本研究建置了一組2,077筆中文醫療問題的多答案摘要資料集Mednet-MAS，每筆資料包含多筆相關答案及人工標記的多答案摘要。英文實驗資料來自MEDIQA 2021國際競賽的MAS資料集，包含286筆英文醫療問題以及多答案的摘要。藉由實驗結果與模型分析得知，我們提出的HCA模型在醫療問題多答案摘要任務達到最好ROUGE和BERTScore分數，比相關研究模型 (BART、PEGASUS、CPT、T5、LongT5、HED) 等有更好的摘要效能，人工評估進一步驗證我們提出的HCA模型在多文檔摘要上有良好的表現。 ;This study focuses on a multi-answer summarization of medical questions, which will benefit the general public by obtaining reliable medical counseling suggestions from relevant answers returned by the question-answering systems. We propose a Hierarchical Clustering with Attention (HCA) model for medical multi-answer summarization. First, we use rhetorical structure theory to extract elementary discourse units and perform hierarchical clustering to group them. A large language model is then used to identify a topic for each cluster. Finally, we integrate discourse units with topic information in each cluster to feed into our proposed hierarchical encoder-decoder architecture, including hierarchical self-attention and cross-attentions to enhance the summary generation performance. Due to the lack of publicly available datasets on Chinese medical multi-document summarization, we manually created a Mednet-MAS dataset, including 2,077 Chinese medical questions with multi-answer summaries, for model performance evaluation. We conducted the experiments on the public MEDIQA-MAS datasets in English and our constructed Mednet-MAS in Chinese. Experimental results indicate that our HCA model outperformed other models like BART, PEGASUS, CPT, T5, LongT5, and HED regarding ROUGE and BERTScore metrics. In addition, human evaluation on randomly selected samples also confirmed our HCA model performs better than the state-of-the-art HED model for medical multi-answer summarization.
