潛在狄利克雷分布技術於文本主題分析之應用：以三位自殺死亡個案的協談內容為例

Thematic Analysis Using Latent Dirichlet Allocation on Suicide Hotline Conversations from Three Callers Who Died by Suicide

主題分析可說是最廣泛應用的質性資料分析方法，但為確保信度和嚴謹性，研究者在過程中必須反覆討論與檢核，卻仍無法避免研究者偏誤。為此，我們利用機器學習演算法潛在狄利克雷分布（latent Dirichlet allocation, LDA）技術和改進了段落—主題命名程序於文本主題分析，再以一位單次自殺企圖個案、二位重複自殺企圖個案生前於自殺防治熱線的協談內容為分析範例，具體說明研究效度和應用方式。首先，本研究比較了現象學質性方法和LDA技術搭配段落—主題命名程序於單次自殺企圖個案協談資料的分析結果，顯示二者的主題命名並不直接等同，但是主題內涵有其一致性，而且保留了語句情境脈絡的段落—主題命名程序似乎更能反映主題的指涉意涵。接著，本研究分析了重複自殺企圖個案於協談中各主題出現的機率和情緒困擾、自殺風險之間的關聯，顯示LDA技術生成主題反映個案主觀的心理構念，不僅呼應協談歷程中的重要心理轉折，更捕捉到自殺死亡前的關鍵心理變化。我們認為LDA技術及段落—主題命名程序具備可接受的信度和效度，而且可搭配其他評量指標或研究工具進行延伸研究，有助益於文本主題分析的知識產出。

Thematic analysis remains the most widely used qualitative analytical method, yet researcher bias and reliability concerns persist despite rigorous validation procedures. This study applied latent Dirichlet allocation (LDA) and an enhanced paragraph-theme naming procedure to address these limitations. We analyzed conversational content from three suicide hotline callers who subsequently died by suicide: one caller with a single suicide attempt and two callers with multiple attempts. First, we compared themes generated by the phenomenological analysis with those produced by LDA and the paragraph-theme naming procedure using data from the first caller. Results indicated that content remained highly similar, although theme labels differed. The paragraph-theme naming procedure appeared superior in preserving contextual nuances and capturing underlying thematic meanings. Second, we examined relationships between theme occurrence probability, emotional distress, and suicide risk in the other two callers. Results demonstrated that LDA-generated themes effectively captured the subjective psychological constructs of the callers. The themes not only echo important psychological turning points during the conversations, but also capture crucial psychological changes preceding suicide death. The LDA technique combined with the paragraph-theme naming procedure demonstrated adequate reliability and validity while offering potential for integration with other assessment tools. This approach represents a significant methodological advancement for qualitative thematic analysis.