- 學位論文
透過起訴書輔助法院判決－以竊盜罪為例
Using Indictments to Assist Judges in Judging – A Case Study with Offenses of Larceny
本文將於2026/02/02開放下載。若您希望在開放下載時收到通知，可將文章加入收藏。
摘要
近年來，隨著技術的成長，自然語言處理的工作在不同的領域間發展，其中亦包含法律面向。在台灣，法律與科技的應用目前仍在起步的階段，有些社群活動亦開始著重於此面向，例如法律科技黑客松。 就台灣的刑事訴訟而言，案件會先經由檢察官的偵查，若被告遭受起訴處分，案件才會移交由法官進行審理及判決。而訴訟的過程往往曠日廢時，其潛在原因可能是被告對於判決結果的不符而上訴。此外，因應國民法官的推動，台灣可能逐步走向參審制的判決。相較於現任法官，國民法官可能沒有法律相關的知識或判決相關的經驗，使其對於最終判決的影響可能較不客觀。 因此本實驗以輔助判決為目標，其對象可以是一般民眾、被告、國民法官，甚至是現任法官及律師等。實驗結合判決結果預測以及類似案件推薦兩部分工作，除了提供使用者可能的判決結果，亦透過「與預測結果相符」及「與預測結果不符」二類相似案件，提供不同面向的案件做為比較及參考。 在過去判決結果預測的相關實驗中，多是以裁判書作為實驗語料。我們則將訴訟流程往回推一步，採用起訴書作為主要語料，希望能在判決結果確定前就對於案件提供相關輔助功能。而在起訴書數量較少，且判決類別不平均的情況下，判決結果預測的實驗目前最高平均值能達到0.665 的Macro_F1分數，在類似案件推薦的實驗中也確實能透過起訴書內容，找出類似案件。
並列摘要
With the advance of science and technology, the works of natural language processing have been growing in many different fields in recent years. In Taiwan, the applications between law and technology are still in their infancy, while some communities have begun to focus on these aspect, such as Legaltech Hackathon. In terms of criminal proceeding in Taiwan, the case will first be investigated by the prosecutor. If the defendant is charged, the case will be transferred to the judge for trial and judgement. But the judicial proceeding is usually time-consuming, and the reason may be that the prosecutor or defendant appealed against the judgement. Furthermore, with the promotions of citizen judge system, Taiwan may gradually move towards a lay judge system. Compared with professional judges, citizen judges may not have legal knowledge or judgement-related experience, which may lead to the less objective judgements. Therefore, the goal of this experiment is to assist judgements, and its objects can be defendants, citizen judges, and even professional judges and lawyers. Our experiment combines two parts of the work, which are judgement predicting and similar cases recommending. In addition to providing users with possible judgements, the two types of similar cases "consistent with the prediction" and "not consistent with the prediction" are also provided for comparison of different aspects of the case. In the past related experiments, court’s judgements were mostly used as experiment corpus. To provide relevant auxiliary functions for the case before the judgement confirmed, we use the indictments as our main corpus. However, with small amount of indictments and unbalanced judgement types, our judgement predicting can still have 0.665 of macro f-1 score, and the similar cases can indeed be found through the content of the indictment.
參考文獻
[1] C. Cortes and V. Vapnik, “Support-vector networks”, Machine Learning, vol. 20, no. 3, pp. 273-297, 1995.
[2] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”, North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL ’19, vol. 1, pp. 4171-4186, Jun. 2019.
[3] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin, “LIBLINEAR: A Library for Large Linear Classification”, Journal of Machine Learning Research, JMLR, vol. 9, pp. 1871-1874, Aug. 2008. Available: http://www.csie.ntu.edu.tw/~cjlin/liblinear/
[4] Z. S. Harris, “Distributional Structure”, WORD, vol. 10, no. 2-3, pp. 146-162, 1954.
[5] A. Heidarian, M. J. Dinneen, “A Hybrid Geometric Approach for Measuring Similarity Level Among Documents and Document Clustering”, IEEE Second International Conference on Big Data Computing Service and Applications, Mar. 2016.