政治網路口碑的情感分析：語意關連性之觀點

Abstract

情感分析是一種以電腦運算方得知文本中是否包含主觀意見與判斷其意見傾向的資料分析技術，包含字典法與機器學習法兩個取徑。本研究採用字典法取徑，以情緒理論為基礎分析2014年九合一選舉中，台北市長候選人連勝文與柯文哲的網路語料。目的為驗證字典法取徑中語意關聯性於中文語料的有效性與探討文本情境對字詞的情緒指標是否有規則性的影響。 研究包含三部分，首先將語料進行斷詞處理，並與情緒字庫進行比對，得出語料中的情緒字詞。實驗一目的為驗證語意關聯性，根據語意關聯性計算公式，計算每一個情緒字詞在語料中與「連勝文」與「柯文哲」之間的相關性。以網路問卷調查法探討字詞與候選人在語料中的關聯性，是否也存在於人的認知關聯性。實驗二將字詞引發情緒的理論基礎納入情感分析研究，關注於使用者對於不同口碑文情境的情緒反應，以實驗法操弄字詞價性、評論候選人與文本情境三個因子，探討其對於價性反應的效果。 實驗一結果顯示，以相關性分析而言，語意關聯性在字詞與候選人相關性的計算上，與人的認知關聯性沒有顯著相關性。但以比較平均數的概念而言，語意關聯性在認知關聯性中得到驗證。正負向字詞的認知關聯性結果不同，政治傾向會影響正負字詞與候選人的關聯性認知，其以泛綠者的認知關聯性反應較極端。整體而言，所有正向字詞在認知關聯性與柯文哲較相關，本研究推測此認知結果除了受到政治傾向的影響之外，也可能來自於口碑文語料中，柯文哲與正向字詞的高度關聯性。實驗二為探討政治口碑文的文本情境是否會影響情緒字詞引發價性的效果。結果顯示，對於柯文哲的句子，負面情境對於情緒字詞的價性引發效果產生調節作用，正向字詞在負面情境對於價性沒有較正向的影響。對於評論對象為連勝文的句子，在情緒概念化階段發生認知情境的轉換，負面情境轉換為正面情境；正面情境則轉換為負面情境。因此負面情境引發的價性感受與正面情境無顯著差異，在平均值上甚至高於正面情境。本研究最後探討實驗結果對於情感分析技術、政治領域的網路民意探勘與網路意見競選策略的理論與實務上的建議。

Sentiment analysis is a kind of data analysis which can figure out the polarity of opinion in text corpus by computer. It has two approaches, dictionary-based approach and machine learning. This study focus on dictionary-based approach and apply it to the corpus about two controversial candidates in the Taipei mayoral election in 2014. The goals of this study is to verify whether the computing of semantic orientation can apply to Chinese corpus and the other is to explore the moderation effect of context to valence. The study has three important parts. First, the corpus undergo the word segmentation and sift out the sentiment words which is contained in it from all of the sentiment lexicon. The goal of first experiment is to test the validity of semantic orientation in Chinese corpuse. Test whether the semantic relation also occur in human cognition by online survey. The second experiment includes the theory of word and emotion elicitation into sentiment analysis, focusing on the emotion response to different context rules of word-of-mouth. The result of first experiment shows that the relation between sentiment words and the name of candidates does not correlate with cognitive relation based on Pearson correlation test. However, this relation is valid in the case of ANOVA analysis. The semantic relation of either positive or negative words with candidates is affected by political orientation. As a whole, all positive words is related to Candidates Ke Wen Zhe. This bias may due to the political orientation or the high relation with positive words and Ke Wen Zhe in corpus. Moreover, the aim of second experiment is to see whether the sentense context in political word-of-mouth will moderate the causal effect of sentiment words to valence. The result shows that, for sentence of Ke Wen Zhe, the effect of positive words is moderated by negative contex. But for the sentence of Lian Sheng Wen, the transition of cognitive relation occurs in the emotion conceptualization stage. Negative context is recognized as positive and vise versa. Therefore negative context have positive effect on valence. The final part of study will discuss the theoretical and practical implication of the result to sentiment analysis, political opinion mining and campaign strategies for online public opinion.