請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92270
標題: | 結合口語資料與大型語言模型的互動式醫療知識問答系統 Incorporating Spoken Data and Large Language Models into an Interactive Healthcare Question-Answering System |
作者: | 郭宜婷 Yi-Ting Guo |
指導教授: | 傅立成 Li-Chen Fu |
關鍵字: | 健康問答系統,對話代理,口語處理,異質資訊網路,大型語言模型, Healthcare Question-Answering System,Conversational Agent,Spoken Language Processing,Heterogeneous Information Network,Large Language Models, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | COVID-19的流行引發了人們在網際網路上尋求健康資訊的行為 (Health Information Seeking Behavior, HISB) 急遽增加。 因此,透過對話代理向公眾傳遞準確的健康知識變得有其必要性。然而,大多數現有的研究僅專注於使用文字資料進行常見問題 (Frequently Asked Question, FAQ) 檢索,忽略了其他如演講和廣播等有價值的口語資源。此外,近年來大型語言模型 (Large Language Models, LLMs) 的快速發展也提供了一個機會,可以結合它們文本生成的能力來開發問答系統,從而提高互動性並增強用戶參與的意願。
在這項研究中,我們提出了一個互動式醫療知識問答系統,基於由醫療專業人員提供的口語資料進行問答檢索。我們的方法包括開發一個逐字稿校正模組,旨在偵測和修正醫學名詞中的中文拼寫錯誤,以及一個問答對提取模組。後者模組與LLM協作執行,取得的問答對也經過驗證以防止幻覺問題。另一方面,為因應口語資料中醫學實體較少且問答對較短的特性,我們設計了一個異質資訊網路(Heterogeneous Information Network, HIN) 來構建相應的結構化資料。此外,我們引入了「主題」的概念,將資料庫中的問答對進行分類,促進多樣數據資源的整合並有效地檢索用戶查詢的答案。在增強互動性方面,我們納入對話歷史來進行用戶查詢重構並提供建議問題,同時利用LLM產生更人性化的回應。實驗結果表明,我們開發的HIN搜索模組在口語資料檢索方面優於其他方法。使用者研究的結果進一步顯示,LLM的結合成功提升了用戶參與度。因此,我們相信利用這個問答系統不僅可以利用各種類型的數據資源,還可以基於多樣的查詢向民眾提供適當的健康知識,從而促進這些有用醫療知識的傳播。 The COVID-19 pandemic has prompted a surge in health information seeking behavior (HISB) on the Internet. Consequently, it is crucial to disseminate accurate healthcare knowledge to the public through conversational agents. However, most existing works only focus on FAQ (Frequently Asked Question) retrieval using textual data, neglecting valuable spoken resources like speeches and podcasts. Additionally, the rapid development of large language models (LLMs) in recent years offers an opportunity to harness their text generation ability for the development of a question-answering system, thereby increasing interactivity and enhancing users'' willingness to engage. In this study, we propose an interactive healthcare question-answering system based on spoken data provided by medical professionals. Our methodology involves the development of a transcript correction module designed to detect and correct Chinese spelling errors in medical entities as well as a question-answer pairs (QA pairs) extraction module. The latter module is executed in collaboration with an LLM, and the obtained QA pairs are verified to prevent happening of hallucination issues. To address the characteristics of spoken data, which tend to have fewer medical entities and be shorter in QA pairs, we design a Heterogeneous Information Network (HIN) to construct the corresponding structural data. Additionally, we introduce a concept, called ``topic", to categorize QA pairs in the database, facilitating the integration of diverse data resources and efficient retrieval of answers to user queries. To enhance interactivity, we incorporate dialogue history to reformulate user queries and offer recommendations, while leveraging LLM to deliver a more humanized response. Experimental results demonstrate that our developed HIN search module outperforms other related works in information retrieval from spoken data. Results from user studies further indicate that the integration of LLM successfully enhances user engagement. Therefore, we believe that utilizing our developed question-answering system not only can leverage various types of data resources but also can deliver appropriate healthcare knowledge to the public based on versatile queries, thereby facilitating easy dissemination of useful medical knowledge. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92270 |
DOI: | 10.6342/NTU202400511 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf 目前未授權公開取用 | 1.66 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。