藉由情感傾向探勘醫療社群媒體

Abstract

隨著網際網路的快速發展，越來越多使用者利用醫療社群論壇來詢問和蒐集一些醫療相關的資訊以及與他人分享個人的醫療經驗。因此，本論文提出一個研究架構，分析醫療社群論壇裡使用者所產生的內容，以提供病人、照顧病人的人和醫生一些寶貴的資訊及知識。所提出的研究架構分為四部分，首先，我們利用每個問題裡的醫療字來建立它的虛擬文件;接著，我們加入一個權重機制來修改LDA (稱作conLDA)以對所有虛擬文件進行分群，將擁有類似醫療字分佈的虛擬文件分至同一病徵主題(C-topic);然後，對於每個病徵主題及每個討論串進行正、負情緒及生理、心理情緒分析，來了解使用者的想法;最後，我們修改Apriori演算法來找出醫療頻繁樣式及分析各個醫療病徵之間的關係，並建立醫療關聯圖(medical association map)。實驗結果顯示，我們的方法可以有效將相關醫療字及談論相關病徵的問題分在一起以及找出較有主題性的病徵主題。本研究所提出的 情緒分析及醫療關聯圖結果，能提供給病人、照顧病人的人以及醫生作快速的參考。

With the rapid development of the Internet, more and more users utilize health communities (known as a forum) to find health-related information, share their medical stories and experience, or interact with other people in health social media. Therefore, in this thesis, we propose a framework to analyze the user-generated contents in a health community. The proposed framework contains four phases. First, we extract medical terms, including conditions, treatments and symptoms to form a virtual document for each medical question, where each medical question may be followed by several comments, and the question and these comments forms a discussion thread. Second, we modify Latent Dirichlet Allocation (LDA) by adding a weighted scheme, called conLDA, to cluster virtual documents with similar medical term distributions into a conditional topic (C-topic). Third, we analyze sentiments in each C-topic and discussion thread by positive and negative polarities, and physiological and psychological sentiments. Finally, we modify the Apriori algorithm to mine frequently mentioned patterns and association rules for building medical association maps. The experiment results show that conLDA outperforms the original LDA and can cluster relevant medical terms and relevant medical questions together. Thus, the C-topics clustered by conLDA are more thematic than those by the original LDA. The results of sentiment analysis and medical association maps may provide a quick reference and valuable insights for patients, caregivers and doctors.