Predicting the Success of Mediation Requests Using Case Properties and Textual Information for Reducing the Burden on the Court

The success of mediation is affected by many factors, such as the context of the quarrel, the personality of both parties, and the negotiation skill of the mediator, which lead to uncertainty for the work of prediction. This article takes a different approach from that of previous legal prediction research. It analyzes and predicts whether two parties in a dispute can reach an agreement peacefully through the conciliation of mediation. With the inference result, we can know whether mediation is a more practical and time-saving method to solve the dispute. Existing works about legal case prediction mostly focus on prosecution or criminal cases. In this work, we propose a long short-term memory (LSTM)-based framework, called LSTMEnsembler, to predict mediation results by assembling multiple classifiers. Among these classifiers, some are powerful for modeling the numerical and categorical features of case information, for example, XGBoost and LightGBM. Some are effective for dealing with textual data, for example, TextCNN and BERT. The proposed LSTMEnsembler aims to not only combine the effectiveness of different classifiers intelligently but also to capture temporal dependencies from previous cases to boost the performance of mediation prediction. Our experimental results show that our proposed LSTMEnsembler can achieve 85.6% for F-measure on real-world mediation data.