毒品施用緩起訴處分預測模型──自然語言與機器學習在檢察書類文本之應用

A Predictive Model for Prosecution Decisions - An Application of Natural Language and Machine Learning To Drug Prosecutorial Documents

隨著人工智慧（Artificial Intelligence, AI）與大型語言模型（Large LanguageModels, LLMs）之進展，自然語言處理（Natural Language Processing, NLP）技術已廣泛應用於語音識別、客服系統、教育諮詢與程式輔助等領域。然而，檢察書類文本因其專業性與結構不規則，過去少有研究成功導入自然語言模型進行自動判讀與分析。本研究以臺灣地方檢察署9,474筆毒品施用案件之起訴書與緩起訴書為資料來源，結合自然語言模型與機器學習方法，從中自動擷取18項書類特徵，並配合32項區域變項，建立結構化資料集，進而訓練五種分類模型（決策樹、邏輯斯迴歸、貝氏分類器、支持向量機與隨機森林），預測檢察官是否作成緩起訴處分。結果顯示，自動判讀模型之文本標註正確率超過80%；在後續預測階段，邏輯斯迴歸於交叉驗證與測試資料上皆展現穩健效能，其準確率達87%、AUC値達0.936，顯示該模型兼具效度與解釋性，具備實務應用潛力。本研究另透過具可解釋性的決策樹模型進一步揭示檢察官決策邏輯。研究結果顯示，「是否累犯」為首要分類節點，預測力最強。於非累犯、未抗辯個案中，「犯後態度」為關鍵次層節點，態度不佳或不明者通常被預測為起訴，態度良好者則進一步由「完成戒癮次數」決定處分方向。此顯示檢察官對於毒品施用者之處分判斷，並非僅依賴靜態特徵，亦重視動態表現與行為修復意願。本研究提供以自然語言模型為基礎之結構化分析途徑，驗證其應用於檢察書類文本之可行性，未來可望作為建置AI初步篩選機制之基礎。在改善檢察機關辦理毒品施用案件之效率上，本研究亦呼籲法務部應於地端建置封閉式AI模型，輔助處理單純毒品施用案件之預篩流程，減輕檢察人力負擔，並推動數據導向之司法判斷透明化與效率化。

With the advances of Artificial Intelligence (AI) and Large Language Models (LLMs), Natural Language Processing (NLP) technologies are used in many fields, such as speech recognition, customer service, educational counseling and programming support. However, due to the professional nature and irregular structure of prosecutorial documents, few studies have successfully applied NLP models to automatically interpret and analyze these documents. This study utilizes a dataset of 9,474 prosecution documents–consisting of indictments and deferred prosecution decisions in drug use cases from district attorneys' offices in Taiwan. By integrating NLP with machine learning, we automatically extracted 18 document-level features and 32 region-level variables to create a structured dataset. Five classification models–decision tree, logistic regression, naïve Bayes, support vector machine, and random forest–were trained to predict whether a prosecutor would make a deferred prosecution decision. The results show that the automatic annotation model achieves a text labeling accuracy of over 80%. In the prediction phase, the logistic regression showed robust and stable performance on both the cross-validation and test sets, achieving 87% accuracy and an AUC value of 0.936, indicating both predictive validity and interpretability. In addition, an interpretable decision tree model was used to reveal the prosecution's decision logic. The results show that“recidivism status”serves as the primary classification node with the strongest predictive power. For non-recidivist defendants who did not contest the charge, 'post-offense attitude' was found to be an important secondary node–those with a poor or unclear attitude were most likely to be predicted to be charged, while those with a good attitude were further assessed by 'number of completed detoxifications'. These results suggest that prosecutorial decisions regarding drug offenders are not only based on static case characteristics, but also take into account dynamic behavioral indicators and the defendant's readiness for rehabilitation. This study demonstrates the feasibility of using NLP-based structured analysis for prosecutorial interpretation of documents and lays the foundation for the development of AI-assisted pre-screening mechanisms. To improve efficiency in processing drug use cases, we recommend that the Ministry of Justice consider implementing a secure AI model in the field to assist in pre-screening routine drug offenses, thereby reducing prosecutors' workload and promoting transparency and data-driven decision making in the criminal justice system.