半導體產業為全球科技發展之基礎，其股價走勢除受傳統量化指標影響外，亦高 度敏感於政策變動與新聞輿情。然而，現有文獻在將多模態融合技術應用於特定高科 技產業時，仍缺乏針對半導體供應鏈的專門研究，且鮮少在同一實驗框架下比較不同 層級自然語言處理技術的實際貢獻差異。 本研究以美國五大半導體龍頭企業(NVIDIA、AMD、Intel、Qualcomm、Broadcom) 為實證對象，建構結合新聞文本情緒與股價時間序列之多模態預測架構。研究設計涵 蓋十種模型組合、三種漲跌幅門檻與五種訓練比例，共產生約一千八百筆實驗紀錄。 文本特徵分別採用 Word2Vec 靜態詞向量與 FinBERT 金融專用預訓練模型提取，並以 雙分支深度神經網路搭配維度瓶頸層及注意力閘門機制進行特徵級融合。 實驗結果顯示，純文本驅動模型之預測表現已可接近傳統技術指標水準，證實新 聞情緒在半導體產業確實攜帶具預測價值之資訊。多模態融合模型在 AUC 指標上於 十五個實驗格位中勝出十四個，AUC 領先比例達 93.3%(14/15 格位)，排序能力之 優勢具跨門檻與跨訓練比例之一致性。惟 F1-Score 於類別不平衡加劇時受統計壓制， 與 AUC 之間呈現系統性剪刀差，據此本研究提出分層指標策略作為評估建議。 FinBERT 與 Word2Vec 之效能差異則呈現場景依賴性，前者在融合架構中表現較優， 後者在單模態場景下更具成本效益。;The semiconductor industry is fundamental to global technological development， and its stock prices are influenced not only by quantitative indicators but also by policy changes and news sentiment. However， existing literature lacks dedicated multimodal studies targeting the semiconductor supply chain， and few works have compared different levels of NLP techniques within a unified experimental framework. This study selects five major U.S. semiconductor companies—NVIDIA， AMD， Intel， Qualcomm， and Broadcom—as empirical subjects， constructing a multimodal architecture that integrates news sentiment with stock price time series. The experimental design covers ten model configurations， three price movement thresholds， and five training ratios， producing approximately 1，800 records. Textual features are extracted via Word2Vec and FinBERT， respectively， and fused through a dual-branch neural network with a dimensionality bottleneck layer and a gated attention mechanism. Results show that text-only models approach the prediction performance of traditional technical indicators， confirming the predictive value of news sentiment in this sector. Multimodal fusion models outperform the baseline under AUC in fourteen of fifteen settings (93.3%)， with consistent advantages across thresholds and training ratios. F1- Score， however， is suppressed under class imbalance， diverging systematically from AUC—a phenomenon this study terms the "scissors effect"—prompting the proposal of a stratified evaluation strategy. The relative performance of FinBERT and Word2Vec proves context-dependent: FinBERT excels within the fusion architecture， while Word2Vec offers greater cost-effectiveness in single-modality settings.
