在數位經濟與工業 4.0 的發展趨勢下，人工智慧（Artificial Intelligence， AI）已成為推動製造業轉型升級之關鍵技術。然而，既有文獻多仰賴主觀問卷、個案研究或單一文字指標衡量企業 AI 導入情形，較難反映企業在不同應用面向之異質性，亦較少從企業前期財務與資源條件探討其對後續 AI 導入揭露之預測關係。 本研究以台灣上市櫃製造業為研究對象，整合公開資訊觀測站（MOPS）年報文本與台灣經濟新報（TEJ）財務資料，透過文字探勘、關鍵字分類與語境篩選規則建構企業 AI 導入揭露指標，並將企業 AI 導入揭露區分為「策略型」、「製程型」與「資訊型」三類應用導向。在實證設計上，本研究採用時間落差（time-lag）設計，以前期財務與企業特徵預測後期年報 AI 導入揭露，並比較 Logistic Regression、Decision Tree、Random Forest、Gradient Boosting、Extra Trees Classifier 、Support Vector Machine與Neural Network等機器學習模型之預測表現。 實證結果顯示，產業特性與企業資源條件共同影響後續 AI 導入揭露；排除產業別後，前期財務與企業特徵仍具邊際預測訊號，其中 Top 20 特徵與設定類別權重之 Extra Trees Classifier 取得相對較佳且穩定之表現。進一步比較顯示，策略型、製程型與資訊型 AI 導入揭露具有不同預測特性，顯示企業 AI 導入揭露並非單一總量概念，而具明顯應用面向差異；敏感度分析亦顯示，不同門檻會影響分類特性，主要門檻 count2 在高導入樣本辨識與整體分類品質間取得較佳折衷。 此外，SHAP 可解釋性分析結果顯示，產業別、企業規模、研發投入、資源條件與資產規模，是模型判斷高 AI 導入揭露企業時的重要特徵。整體而言，本研究透過年報文本與財務資料之整合，建立企業 AI 導入揭露之預測架構，並以多維度 AI 導入揭露型態與可解釋機器學習方法，補充既有研究對 AI 導入衡量與預測分析之不足。研究結果可作為後續探討企業 AI 導入揭露、數位轉型與財務資源條件關係之實證參考。;Under the development trends of the digital economy and Industry 4.0， artificial intelligence (AI) has become a key technology driving the transformation and upgrading of the manufacturing sector. However， prior studies have often relied on subjective surveys， case studies， or single textual indicators to measure firms’ AI adoption， making it difficult to capture heterogeneity across different application dimensions. Moreover， relatively few studies have examined the predictive relationship between firms’ prior financial and resource conditions and their subsequent AI adoption disclosure. This study focuses on publicly listed and over-the-counter manufacturing firms in Taiwan. It integrates annual report texts from the Market Observation Post System (MOPS) with financial data from the Taiwan Economic Journal (TEJ). Through text mining， keyword classification， and contextual filtering rules， this study constructs firm-level AI adoption disclosure indicators and further classifies AI adoption disclosure into three application orientations: strategic， process-oriented， and information-oriented. In terms of empirical design， this study adopts a time-lag approach， using prior-period financial and firm characteristics to predict subsequent AI adoption disclosure in annual reports. It also compares the predictive performance of several machine learning models， including Logistic Regression， Decision Tree， Random Forest， Gradient Boosting， Extra Trees Classifier， Support Vector Machine， and Neural Network. The empirical results show that industry characteristics and firm resource conditions jointly influence subsequent AI adoption disclosure. After excluding industry information， prior financial and firm characteristics still retain marginal predictive signals. Among the models， the Extra Trees Classifier with Top 20 features and class-weight adjustment achieves relatively stable and favorable performance in the overall AI adoption disclosure prediction task. Further analysis of different AI adoption disclosure types shows that strategic， process-oriented， and information-oriented AI adoption disclosure exhibit distinct predictive characteristics. This finding suggests that firms’ AI adoption disclosure is not a single aggregate concept， but rather contains clear differences across application dimensions. Sensitivity analysis also indicates that different AI adoption disclosure thresholds affect classification performance， with the main threshold， count2， achieving a relatively favorable balance between identifying high-disclosure samples and maintaining overall classification quality. In addition， the SHAP explainability analysis shows that industry category， firm size， R&D investment， financing conditions， and asset scale are important features for the model when identifying firms with high AI adoption disclosure. Overall， by integrating annual report texts and financial data， this study establishes a predictive framework for firm-level AI adoption disclosure. Through multidimensional AI adoption disclosure indicators and explainable machine learning methods， this study complements prior research on the measurement and prediction of AI adoption. The findings provide empirical references for future studies on the relationships among AI adoption disclosure， digital transformation， and firms’ financial resource conditions.
