運用文字探勘分析全民健保與五項總額支付制度之民眾知覺感受

Application of text mining in the public perception analysis of global budget payment and National Health Insurance systems

目標：本研究為發展與評估一套健保文字探勘技術以彌補傳統量性監測工具的不足並克服質性資料分析與解讀上的困難。方法：以106年度全民健保與五項總額支付制度之全國民意調查為基礎。運用文字採礦技術探討民眾對於健保意見感受之語意結構。進一步將上述語意結構進行情感分類，評估其做為輔助監測民意之適切性。結果：六大文本的詞頻統計皆近似齊夫經驗法則，少數高頻詞彙影響多數的文本內容。透過文字雲得到共有語意特徵為「很好」、「滿意」和「不錯」等對健保正向的評價。於因素分析皆呈現多元的潛在語意結構，整體解釋變異達70％以上。情感標記圖之分布皆呈反拋物線，點座落位置與色彩間的對比差異說明質、量性資料彼此具互補效益。結論：本研究具體實證如何從公眾的自由評論中，提取出輿情感受與意向，同時藉由情感標記，偵測文字評論背後的情感與情緒，將可作為探討民眾對於健康政策感知的另一項調查方法依據。 Objectives: The aim of this study was to facilitate text mining for overcoming the limitations of traditional questionnaire-based surveys and the challenges of analyzing textual data for monitoring public perceptions regarding the different fields of the National Health Insurance (NHI) system. Methods: The study data consisted of the 2017 national survey and public polls regarding people's satisfaction with NHI and five global budget payment systems in Taiwan. We derived the structure and information from the text of responses through text mining and further applied sentiment analysis by using lexical signifiers and classifying emotions to determine whether the data lean toward positive or negative emotions. Results: (1) The rank-frequency distribution for all texts of the six major fields of medical care followed Zipf's law. This means that a few high-frequency words affected most of the textual content. (2) The results in word cloud visualizations could reflect public perceptions regarding various topics. The most common semantic features were ＂very well,＂ ＂satisfied,＂ and ＂good.＂ Most participants, especially dialysis patients, provided positive comments on different types of medical care in the NHI system, which shows that the health care offered by the NHI system meets peoples' needs. (3) Factor analysis further indicated that our semantic results explained more than 70% of the variance for the six major texts. For the majority of the dialysis patients, the semantic structure of responses consisted of positive words and was significantly different from those of other texts. (4) The reverse para-curve and color differences in valence-arousal space further produced qualitative and quantitative results indicating the subjects' positive perceptions and opinions. Conclusions: Public perceptions and opinions are essential for policy evaluation and implementation. They play a critical role in policy innovations. This study outlines the potential benefits of text mining and sentiment analysis techniques for developing an NHI sentiment analysis system that can aid in the assessment of public perceptions and opinions regarding the health care system. Police surveillance can benefit by incorporating such a systematic approach into the visualization and standardization of data content.