真實生活中人們遇到醫療問題，經常藉由不同的管道，尋求醫生的建議與解答，而自動問答系統提供一個即時回覆答案的解決方案。本研究的主要目標為建立中文醫療問答系統，將問題輸入問答系統，從醫療問答資料集中，匹配找出最佳的答案返回給使用者。近年來，不同於傳統的詞彙匹配，深度學習的興起帶動了語義匹配的方式，深度語言模型能有效學習文本的語義訊息，並藉此找出相近的文本。許多研究均顯示出語義匹配的方法較傳統的方法得到更好的效果，因此，我們提出句嵌入向量重排序器 (Sentence Embedding Reranker, SER) 模型。 中文問答資料來自於醫聯網 (https://med-net.com/)，資料集共有 26,816 筆醫療問答，我們使用 Pooling method 建立系統測試集，從 26,816 筆問題中取 120 筆問題作為測試問題，每個問題分別經過兩個不同的檢索系統 (BM25 以及 Sentence-BERT)，返回100 筆答案，並人工標註其答案的正確性，最後取兩系統的聯集作為系統測試集。藉由實驗結果得知，我們提出的 SER 重排序器模型，在 MAP、NDCG 效能指標達到最好的分數，有效增進中文問答系統的檢索效能。 ;In the digital era, users usually search and browse web content to obtain healthcare related information before making a doctor’s appointment for diagnosis and treatment. The automatic question-answering system can provide a solution to address this need in real-time. Our main research objective is to design and implement a Chinese medical question answering system. In such a medical QA system, users issue a question as a query and then obtain relevant doctors’ answers in the ranked list. Different from traditional lexical matching methods, the deep learning-based semantic matching model can effectively learn the semantic features to retrieve similar texts. Therefore, we propose a Sentence Embedding Reranker (SER) model to enhance the question-answering performance. The Pooling method was used to combine the top 100 results returned by BM25 and Sentence-BERT retrieve systems for answer relevance annotation. Based on experimental results from these manual-annotated question-answer pairs, our proposed SER re-ranking model achieved the best results in MAP and NDCG, which can enhance the performance of the Chinese medical question-answering system.
