植基於文字探勘的立委候選人選舉政見相似度分析

林玗靜 ； 戴敏育 ； 汪志堅 ； 陳建彰

《TANET2017 臺灣網際網路研討會》 2017年 (2017/10) Pp. 991-996

https://doi.org/10.6728/TANET.201710.0172

摘要

台灣辦理公開選舉多年，多年來一直有一種說法，主張台灣選民不會參考候選人政見，而只會根據政黨傾向或候選人之選區經營狀況來決定投票行為。這種說法是否為真，並無明確的證據。本研究分析2016年立委選舉資料，利用文字探勘的N-Gram斷詞法建立政見詞庫，並據以進行詞頻分析。利用相似度分析，計算當選與非當選候選人政見的相似性，並利用統計次數分佈計算依政黨、區域、性別、選舉結果、身份別等五種因素之下的候選人政見方向。本研究結果顯示，各個當選候選人的政見的平均相似度，高於當選和落選候選人政見間的相似度，而落選候選人間政見平均相似度最低。詞頻分析之實驗結果顯示，都會區與非都會區、不同政黨、是否為原住民、當選與否等不同候選人，在政見內容上有顯著差異。此一結果顯示各種不同身份候選人的政見並非全然相似，且政見也與當選與否相關。

關鍵字

選舉政見 ； 詞頻分析 ； N-Gram ； 文字相似度 ； 文字探勘

並列摘要

Elections are held in Taiwan since 1946. Over the past few decades, people sometimes argue that Taiwanese voters usually vote based on the candidate's political party or constituency service, rather than policies. However, no evidences are provided for this argument. The current study analyzes the bulletin of 2016 Taiwan legislator election to acquire the relationships between candidate and their policies. By using the N-Gram text mining technique, we established the policy domain dictionary. We analyzed the similarities of policies of candidates based on their parties, regions, genders, election results and identity of aborigines. The results show that the average similarity of polities among elected candidates are highest, when compare with similarity of politics between elected and unsuccessful candidates, and group of politics among unsuccessful candidates are the lowest. The result of word frequency showing: when candidates grouping according to their parties, regions, election results and identity of aborigines, their policies have significant difference.

並列關鍵字

policies ； word frequency analysis ； NGram ； word similarity ； text mining