| 摘要: | 此近幾年運用機器學習（Machine learning ）與深度學習（Deep learning）方式對於繁體中文、簡體中文、英文等法律文本進行法律專業術語的標記、法律罪章名稱的分類以及法律量刑範圍的預測逐漸為人們所關注，由此國內因國民法官的議題而也有了AI法官議題的延伸，希冀透過更理性、客觀與一制性的邏輯處理能力準確一致的進行案件的裁量。其中據OpenAI實驗室的測試Generative Pre-trained Transformer (GPT)模型如果參加美國紐約州（State of New York）所舉辦的律師司法考試，成績將可落在平均值前10%的高水準，具有相當優秀的法律知識成果，見其Transformer模型於司法文書的應用上的成果。將中文法律領域於機器學習與深度學習的應用，最先碰到的是法律資料結構化的問題，中華民國法律在資料結構化方面處理不易。由於法律資料的數量龐大且法條種類繁多，資料之間的關係和層次復雜，因此需要進行有效的資料結構化和組織化處理。本次AI法官模型研究中的模型架構分為資料前處理、AI法官系統裁量特徵選取、模型訓練與預測三個階段，資料前處理階段會自司法院資料開放平臺下載全國各級法院1996年1月份至2022年8月份裁判書資料JSON開放資料格式檔案。以Bidirectional Encoder Representati-ons from Transformers進行裁判書文本的處理，可通過對法律裁判書的文本來進行訓練，學習到裁判書中的語意信息和詞彙的關係，並將裁判書內容轉換為向量表示，取出資料集中酒後駕車案件類型的特徵。最後以多任務學習multi-task learning結合自動調整loss weight的方式進行模型的訓練與預測，以同時得到罪責及刑期兩種預測結果，準確率可達到0.95。;In recent years, the use of machine learning and deep learning methods for tagging legal terminology, classifying legal chapter names, and predicting legal sentencing ranges in legal texts written in Traditional Chinese, Simplified Chinese, and Englis- h has gradually become a topic of interest. This has led to the extension of the AI judge issue in domestic circles due to the issue of citizen judges. The hope is to us- e more rational, objective, and standardized logic processing capabilities to make a- ccurate and consistent judgments in legal cases.According to tests conducted by O- penAI Laboratory, the Generative Pre-trained Transformer (GPT) model can achie- ve high-level scores, placing it in the top 10% on the New York State Bar Exam. This indicates that the model has achieved excellent legal knowledge results, dem- onstrating its effectiveness in the application of transformer models in legal doc- uments.In the application of machine learning and deep learning in the Chinese le- gal field, the first challenge encountered is the problem of legal data structuring. The legal system in the Republic of China (Taiwan) is not easy to handle in terms of data structuring. Due to the vast amount of legal data and the complexity of the relationships and hierarchies between different types of legal provisions, effective data structuring and organization is necessary. In this study of the AI judge model, the model architecture is divided into three stages: data preprocessing, feature sele- ction for the AI judge system′s discretion, and model training and prediction. In the data preprocessing stage, the national court′s judgment data in JSON open data format from January 1996 to August 2022 is downloaded from the Judicial Depart- met Open Data Platform. The Bidirectional Encoder Representations from Trans- formers (BERT) is used to process the text of the judgments. By training on the text of legal judgments, the model learns the semantic information and vocabulary relationships within the judgments and converts the judgment contents into vector representations. Relevant features are extracted from the dataset for cases involve- ng drunk driving. Finally, the model is trained and predicted using multi-task lear- ning and automatic loss weight adjustment to obtain both the prediction results for the criminal liability and the sentence length, with an accuracy rate of 0.95. |
