文本情緒推論：封閉式與開放式詞彙分析比較

Sentimental Analysis: Comparison between Closed and Open Vocabulary Analyses

封閉式詞彙分析（closed vocabulary analysis）與開放式詞彙分析（open vocabulary analysis）是兩種電腦化文本分析（computerized text analysis）取徑。前者主要依據一套依心理語文屬性編列的詞典，記算各詞類（例如，正向情緒詞）於文本中出現的頻次。心理學家發現這些詞類的使用確實能反映書寫者的內在狀態。後者則是使用機器學習演算法或人工智慧（artificial intelligence, AI）語言模型，直接針對文本抽取語文特徵，例如，以AI語言模型針對文本詞彙生成相應的語意表徵。過去研究發現開放式詞彙抽取出的語文特徵比封閉式詞彙分析中的詞類，更能正確預測書寫者的性別年齡等變項。然而，這些人口變項往往對應了特定詞彙的使用，但這些詞彙卻可能不曾被收錄於封閉式詞彙分析的詞典中，如此有可能只有利於開放式詞彙分析。同時，這些研究並沒有比較過AI語言模型，以及它們都是以線性迴歸模型檢驗語文特徵預測依變項的正確率。基於上述幾點，本研究改以書寫方式較為豐富多變的情緒文本為刺激材料，使用三種封閉式詞彙分析詞典與一個AI語言模型Bidirectional Encoder Representations from Transformers（BERT），對這些文本抽取語文特徵。然後訓練一個三層神經網路（非線性模型）根據這些語文特徵預測文本的情緒價性。結果發現，使用BERT能更正確地預測測驗文本的情緒價性，這應該是因為BERT生成的語意表徵也能表徵語徑脈絡所致。

Closed- and open-vocabulary analysis are two approaches to computerized text analysis. In closed-vocabulary analysis, the frequencies of words in a text that have been collected in a psycholinguistic dictionary are counted. Psychologists have indicated that the frequencies of words in such a dictionary reflect mental status. In open-vocabulary analysis, machine learning or artificial intelligence (AI) are used to extract textual features. AI language models can generate semantic representation vectors for words. According to previous studies, the features extracted in open-vocabulary analysis can accurately predict the authors’sex and age, among other attributes. However, this finding requires reexamination because flaws in the studies may have been overlooked. First, these demographic variables may have been marked by the cohorts of specific words, which may not have been collected in the dictionary, thereby favoring open-vocabulary analysis. Second, none of these studies involved an AI language model. AI language models are developed for general natural language processing and are therefore suitable for vocabulary analysis. Third, the performance of closed- and open-vocabulary analysis has been evaluated using linear regression, with extracted features included as predictors and demographic variables included as dependent variables. However, no linear relationship has been identified between linguistic features and dependent variables. In this study, we compared closed-and open-vocabulary analysis. For closed-vocabulary analysis, three dictionaries were tested, and for open-vocabulary analysis, an AI language model called BERT was tested. In the closed-vocabulary analysis, words in dictionaries were used as linguistic features, whereas in the open-vocabulary analysis, tokenized representation vectors for words in a text were used as linguistic features. These linguistic features were fed into a three-layered neural network to identify linear and nonlinear relationships between the independent (input) and dependent (output) variables. The output nodes of this model were subsequently corrected to understand the sentiment of the text (i.e., positive or negative). According to the results, BERT outperformed the three dictionaries in terms of predicting the sentiment of the texts, a finding consistent with those of previous studies. This high performance was presumably because the linguistic features generated by BERT represented both the meanings of the words and their relationships with neighboring words (i.e., context).