法官判決時的公正性被大眾所懷疑時有所聞，尤其在重大社會案件發生時面對輿論壓力，即使拿出許多法源作出解釋，依然會有許多質疑聲浪。近年來人工智慧日新月異，在自然語言處理的領域也有許多研究，因此，我們利用臺灣各層級法院判決書資料，擷取判決書文章中僅「事實及理由」的部分，進行分析並預測被告刑期，法官可用此證明其在相似案例上的見解與其他法官相同，或公眾也可以藉此監督法院的判決是否公正，希望能據此構建法院與公眾之間一條透明的溝通管道，達到輔助法院並造福民眾的目的。
本文提出一個根據案件事實描述，便可以預測被告刑期的模型。我們先使用基於規則的字串處理過程篩選並清理資料，將繁體中文的文字做斷詞處理，接著利用預訓練模型將詞語轉為電腦看得懂的數字向量形式，並使用合成少數類過採樣技術(Synthetic Minority Oversampling Technique, SMOTE)處理資料不平衡的問題，最後使用深度學習模型做訓練與學習。
我們效能最好的過取樣加上修改架構後的文本分類卷積神經網路模型(SMOTE + TextCNN)，擁有可以觀察前後文資訊的能力，在政府公開的公共危險案件判決書資料集上，無容忍刑度誤差之準確率可達到73.96%，可容忍刑度誤差前後一個月之準確率可達到95.60%，並在執行速度上每秒可做1,524次判決，亦即達到可以非常快速地做出判決。此外，我們也利用注意力機制加上雙向門控循環神經網路模型(Bi-GRU)，達成模型學習可視化的效果，拆解神經網路時常被人詬病的黑盒子(Black-Box)式的運算學習過程，藉此可以看到神經網路在做出決策時，於判決書文章中所注意的單詞。

Sometimes the justice of the judge judgment is suspected by the public at large, especially when facing with the pressure of the public at large opinion if a major criminal case occurs. Even if many legal sources are used to explain, there will still be many voices of doubt. In recent years, artificial intelligence has changed rapidly, and there have been many studies in the field of natural language processing. Therefore, we use the information of Taiwan court judgments at all levels to extract only “Facts and Ratio decidendi” in the judgment article, analyze and predict the legal penalty of the defendant. The judge can refer the above legal penalty to prove that his views on similar cases are the same as those of other judges, or the public at large can also supervise whether the court judgment is fair. According to this, we hope to build a transparent communication channel between the court and the public at large, and achieve the purpose of assisting the court and benefiting the people
This thesis proposes a model that can predict the legal penalty of the defendant based on the facts of the case. We first use a rule-based string processing process to filter and clean up the data, use traditional Chinese characters for word segmentation, and then apply a pre-trained model to convert the words into a digital vector form that the computer can understand. Next, we adopt the Synthetic Minority Oversampling Technique algorithm to deal with the problem of unbalanced data, and finally employ a deep learning model for training and learning.
Our best performance oversampling combined with a modified Text Classification Using Convolutional Neural Network (SMOTE + TextCNN) has the ability to observe contextual information. On the Offenses against public safety case judgment dataset, the accuracy without tolerance error can reach 73.96%, while the accuracy with more and less one-month can reach 95.60%. As for the execution time, the speed can make 1,524 judgments per second. It means that our legal penalty prediction model has the ability to make judgments in real time.
In addition, we also use the attention mechanism and the bidirectional gated recurrent neural network model (Bi-GRU) to achieve the effect of model learning visualization that makes the Black-Box of the neural network interpretable. In consequence, you can see the words that the neural network notices in the judgment article when making decisions.

[1] 中華民國法務部, 中華民國刑事訴訟法第155條, (2019). Accessed: Sep. 8, 2019. [Online]. Available: https://law.moj.gov.tw/LawClass/LawAll.aspx?pcode=C0010001
[2] U. J. Schild, “Criminal sentencing and intelligent decision support,” in Proceedings of the Judicial Applications of Artificial Intelligence, Dordrecht, Netherlands, pp. 47-98, 1998.
[3] J. Kolodner, Case-Based Reasoning, San Mateo, California, USA: Morgan Kaufmann, 2014.
[4] M. J. Theresa and V. J. Raj, “Analogy making in criminal law with neural network,” in Proceedings of the International Conference on Emerging Trends in Electrical and Computer Technology. Nagercoil, India, pp. 772-775, 2011.
[5] W. C. Lin et al., “Exploiting machine learning models for Chinese legal documents labeling, case classification, and sentencing prediction,” International Journal of Computational Linguistics and Chinese Language Processing, vol. 17, no. 4, pp. 49-68, 2012.
[6] 中華民國司法院, 量刑趨勢建議系統, (2018). Accessed: May. 20, 2020. [Online]. Available: https://sen.judicial.gov.tw/pub_platform/sugg/index.html.
[7] J. Han et al., “Analysis of criminal case judgment documents based on deep learning,” in Proceedings of the International Conference on Advanced Control, Automation and Artificial Intelligence, Shenzhen, China, pp. 230-233, 2018.
[8] H. Zhong et al., “Legal judgment prediction via topological learning,” in Proceedings of the International Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, pp. 3540-3549, 2018.
[9] J. Sun, “Jieba”, (2012). Accessed: Sep. 20, 2019. [Online]. Available: https://github.com/fxsjy/jieba.
[10] G. E. Hinton, “Learning distributed representations of concepts,” in Proceedings of the Eighth Annual Conference of the Cognitive Science Society, Amherst, Massachusetts, pp. 1-12, 1986.
[11] T. Mikolov et al., “Efficient estimation of word representations in vector space,” 2013, arXiv:1301.378.
[12] N. V. Chawla et al., “SMOTE: Synthetic minority over-sampling technique,” Journal of Artificial Intelligence Research, vol. 16, no. 1, pp. 321-357, 2002.
[13] D. E. Rumelhart et al., “Learning representations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533-536, 1986
[14] V. Nair and G. E. Hinton, “Rectified linear units improve restricted Boltzmann machines,” in Proceedings of the 27th International Conference on Machine Learning, Haifa, Israel, pp. 807-814, 2010.
[15] G. E. Hinton et al., “Improving neural networks by preventing co-adaptation of feature detectors,” 2012, arXiv:1207.0580.
[16] K. Cho et al., “Learning phrase representations using RNN encoder-decoder for statistical machine translation,” in Proceedings of the International Conference on Empirical Methods in Natural Language Processing, Doha, Qatar, pp. 1724-1734, 2014.
[17] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735-1780, 1997.
[18] J. Chung et al., “Empirical evaluation of gated recurrent neural networks on sequence modeling,” presented at the Neural Information Processing Systems 2014 Deep Learning and Representation Learning Workshop, Montreal, Canada, Dec. 12, 2014.
[19] M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural networks, ” IEEE transactions on Signal Processing, vol. 45, no. 11, pp. 2673-2681, 1997.
[20] V. Mnih, N. Heess, and A. Graves, “Recurrent models of visual attention,” in Proceedings of the Conference on Advances in Neural Information Processing Systems, Montreal, Canada, pp. 2204-2212, 2014.
[21] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” presented at the 3rd International Conference on Learning Representations, San Diego, California, May 7-9, 2015.
[22] Y. LeCun et al., “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, 1998.
[23] Y. Kim, “Convolutional neural networks for sentence classification,” in Proceedings of the 19th Conference on Empirical Methods in Natural Language Processing, Doha, Qatar, pp. 1746-1751, 2014.
[24] Tensorflow, “Tested build configurations,” (2019). Accessed: Sep. 2, 2019. [Online]. Available: https://tensorflow.google.cn/install/source_windows#tested_build_configurations.
[25] 中華民國司法院, 司法院裁判書公開資料, (2017). Accessed: Sep. 8, 2019. [Online]. Available: http://data.judicial.gov.tw/.
[26] 中華民國法務部, 全國法規資料庫刑法沿革, (2019). Accessed: Sep. 8, 2019. [Online]. Available: https://law.moj.gov.tw/ENG/LawClass/LawHistory.aspx?pcode=C0000001
[27] 中華民國立法院國會圖書館,中華民國刑法第185-3條, (2015). Accessed: Sep. 8, 2019. [Online]. Available: https://lis.ly.gov.tw/lglawc/lawsingle?0^980606A181608C21060681F260609866A681C160621B66A6A181
[28] 中華民國法務部, 中華民國刑法第59條, (2019). Accessed: Sep. 8, 2019. [Online]. Available: https://law.moj.gov.tw/ENG/LawClass/LawParaDeatil.aspx?pcode=C0000001&bp=11