法律資料分析的優化與應用：以離婚後未成年子女親權酌定的裁判為素材

夫妻離婚後，法院如何將具有高度不確定性之「子女最佳利益原則」，貫徹於實際親權酌定個案中，往往是當事人激烈爭議的焦點。儘管有民法第1055條之1提供各項判斷方針，依然難以解決「子女最佳利益原則」判斷標準模糊不明之困難。本研究統計民國98年1月1日起至106年12月31日止全台地方法院有關親權酌定第一審的2,775件裁判，計未成年子女4,340人，計劃透過實證裁判的統計與量化分析，以歸納法試圖尋繹出親權酌定在實務案件的實踐原則。本研究之取樣係以司法院法學資料檢索系統為範圍，而法院設定裁判不公開之作為並非隨機，此將影響研究取樣的隨機性，而成為本論文所有統計量化分析與立論的最大限制。再者，本研究事實的認定以裁判書為準，採法官的觀點，不作事實認定的工作，故而本論文預測模型的預設使用對象為法官、而非律師或當事人，並無取代律師、法官的可能。第三，本研究僅以研究者一人間隔一段時間重新編碼的方式改正失誤，防錯、偵錯、除錯的機制均不夠嚴謹。最後，目前裁判書的編碼工作尚未找到可以避免觀察者錯誤效應的標準流程，可能會因此導致預測模型準確率的高估。本研究羅列了對數機率迴歸、決策樹、隨機森林、梯度型推進決策樹、類神經網路等5種預測工具。發現它們預測裁判結果的準確率差別有限，皆可達95.5%以上，整體而言本研究的資料型態較適合使用梯度型推進決策樹。若使用本研究建議的客製化流程，準確率可再提高，且預測的信度上升。為解決預測模型「解釋性受限」的問題，本研究提出在預測模型之外另行建構論理模型的主張，根據本研究的樣本完成樣式化論理模型－－以親子意願交叉比對、輔以親職能力、現住所等因素所建構的高維複合邏輯路徑模型。不管是整體樣本、或其他分眾樣本，其解釋覆蓋率介於96.93%-99.27%，已能成功擔當實證裁判研究的論理任務。我們以前述5種客製化預測模型逐年滾動分析近五年的裁判資料，發現準確率維持在97.25%-99.86%之間，可見這5種預測模型都具有預測全新未知樣本的能力。雖然進一步科曼哈卡方檢定發現103年的資料因為無書狀母親的事件較多而與其他年份的資料在分布上有顯著的不同，唯不影響本研究預測模型與論理模型的表現。本研究進一步分析與親權酌定相關的28種因素，發現當訴訟一方為外籍配偶、或有一造未到庭且無書狀聲明陳述意見時，其對造都有較高的比率取得未成年子女的親權。在地域的分布上，一造未到庭且無書狀聲明陳述意見的事件比較偏向非六都地區。子女年齡與親權歸屬有關聯性，分別在0-2歲、及14-19歲的兩極出現親權歸母的高峰。本研究在論理模型的導引下，發現這是因為「幼兒從母」、「中兒唯父有意」及「大兒擇母」三股不同力量共同作用的結果，而後兩者都與外籍母親有關。至於其他因素，除子女的排行、父母教育程度與親權歸屬無關之外，主要照顧者等另18個因素的組間比較都達顯著差異。我們利用對數機率迴歸，檢定出權值較大的9個因素依序為：父母意願、子女意願、親職能力、父母品行、支持系統、主要照顧者、現住所、經濟能力、對子女有不當行為。本研究發現，社工的利用率高（95.36%）且98.74%的親權建議與裁判結果相同。而程序監理人、家事調查官則利用率低，分別為1.32%、及1.48%。程序監理人97.62%的建議與裁判結果相同，家事調查官則是100%一致。有關外籍配偶，本研究發現有著城鄉分布的差距：以台籍父母為基準，陸籍配偶的分布偏向六都地區，而越印等其他籍則偏向非六都。個別因素分析結果，不論陸籍或越印其他籍母親組，法官裁判標準其實是一致的－－都是以父母及子女的意願最具影響力。細部檢定發現，台籍、陸籍、越印等其他籍母親出庭有書狀的比率出現由高而低的變化，她們取得親權的比率也是如此趨勢。即便有書狀的外籍母親，其獲得親權的比率還是不如台籍。本研究藉由七維表格分析、樣式化論理模型的樣本流檢定等數據，進一步分析發現外籍母親在親子意願、及主要照顧者、現住所、支持系統都是有利父多於有利母，且達顯著差異，試圖說明為何外籍母親取得親權的比率會低於台籍母親。

關鍵字

法律資料分析；親權酌定；預測模型；論理模型；幼兒從母原則；外籍配偶

並列摘要

Applying Artificial Intelligence on litigation prediction has developed for more than 40 years. A litigation predicting model can play a role of communicator, opening a window for us to look outside. The more accurate it is; the more people will embrace it. Several models have been applied in the field of litigation prediction. Hu and Huang predicted the guardianship for the ward by regression; Huang and Shao operated models of decision tree and artificial neural network to analyze child custody in divorcing cases. The results of their studies show surprising accuracy. In addition, some ensemble learning machines, such as random forests and gradient boost, also lead to widespread attentions. However, the database in Huang and Shao's study was restricted to the parents with equal intention to fight for the custody only, just about one fourth of the whole child custody cases, whereas we need a comprehensive case bank, allowing us to rearrange data to find a better set for the the training of the predicting models. On the basis of the aforesaid database, we can also compare these predicting models to find a better one, and optimize their performance. Furthermore, there might be possibility to evolve a reasoning model from them, which can avoid the potential conflicts arising from the explaining limitation of the predicting models. We collected all the child custody cases from 2009 to 2017 and applied the predicting models as well as the statistical tests for analysis. We find when one of the parents is not a native, or when he or she did not appear in court, the counterpart would have a good chance to be awarded the custody. The families resided in the rural area tended to not appear in court. The ages of the children are correlated with the custody. If they are younger than 2, or older than 14, their mothers would have a tendency to be entitled the custody. By the aid of a templated reasoning model, called "i-like", we find three phenomena responsible: "the tender years doctrine", "the predominance of the father's intention of fighting for the custody when the child is a tween", and "the teenager's preference for mother". The latter two have something to do with the characteristic foreign spouses in Taiwan. When it comes to the other related facors, all but the rankings of the child and the education of the parents are related to custody decision with statistical evidence. The rankings of the factors determined by logistic regression are: the parent's intention, the child's preference, the parent's proficiency, the parent's conduct, the supporting system, the major caregiver, the child's present residence, the parent's economic status, and the happening of undue behavior to the child. As for the foreign spouses, while comparing with the Taiwanese ones, the families having foreign spouses from mainland China were prone to reside in the urban area, whereas the families with ones from those other than mainland China tended to reside in the rural area. The possibility of being rewarded the custody for them both fell far behind the Taiwanese spouses. Possible reasons may be they had a tendency to not appear in court, and they had less competence in the factors of the major caregiver, the child's present residence, and the supporting system from the analysis of a 7-way table. Applying "i-like" model on them, we could also deduce similar reasons closely relative to their nativity by its major factors.