Juan, YiningYiningJuanChen, Chung-ChiChung-ChiChenChen, Hsin-HsiHsin-HsiChenWang, Daw-WeiDaw-WeiWang2026-03-112026-03-112023-11[9798891760165]https://www.scopus.com/record/display.uri?eid=2-s2.0-105027157801&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/736220Predicting child custody decisions post-divorce is crucial but challenging due to numerous nonnumerical, text-based factors, particularly in joint custody scenarios. This study presents the Intermediate Self-Supervised Training (ISST) method, a two-stage approach that classifies document paragraphs using original rationale labels before leveraging this to predict custody at the document level. Achieving up to 90.57% accuracy and notably, a 78.95% F1-score for joint custody cases, it surpasses previous models by 13%. We further refine the model to mitigate gender bias in the training data and provide error estimations, enhancing fairness and reliability. Our user-friendly online system exemplifies our model's applicability in out-of-court dispute resolution, potentially reducing time and financial strains for families in crisis.trueconference paper10.18653/v1/2023.ijcnlp-demo.22-s2.0-105027157801