運用機器學習預測法院裁判──法資訊學之實踐

Predicting Family Court Cases by Machine Learning: An Application of Legal Informatics

電腦（機器）能否模擬人類思考並決策？本文以二○一二年初至二○一四年底共三年期間，地方法院第一審共四百四十八件結果為「單獨親權」之裁判中的六百九十位未成年子女為樣本，將法官考量因素編碼後，使用機器學習領域中的類神經網路（ArtificialNeuralNetwork）方法，訓練機器分析當父母均有意願爭取親權時，何者能獲得法院之支持，並以此為基礎，進而預測其他未知事件的結果。在一百次反覆的操作中，機器之正確率平均達百分之九十八以上。本文之研究成果，有助預測大多數親權酌定裁判的結果（親權歸屬父或母），可供司法從業人員作為判斷之參考，提高裁判之可預測性及法之安定性。

Is it possible that the computer (machine) thinks and makes decisions as human beings? This research collects 448 cases including 690 children in which parents are both Taiwanese and willing to acquire the custody but adjudicated as sole custody as a result by district courts in 2012 to 2014. First, we coded the factors that have been considered by the judge in each case. Second, the artificial neural networks (ANN) were developed using these coded cases and a rigorous training and testing protocol. The machine learned whether the father or the mother receives custody in 552 training cases. Based on the rule learned, the machine then predicted the results of the remaining 138 testing cases that the answers were “unknown” to it. The ANN correctly decided 98% of the cases (accuracy) and the F-1 score reached 0.99 averagely. This result indicates that ANN technology can be used as a beneficial reference for the family judges while deciding child custody. It may also help to raise the predictability of cases as a whole.