基於多BERT模型之NLLP應用於建築工程訴訟之理解與預測

NLLP for the Understanding and Prediction of Construction Litigation Based on Multiple BERT Model

本研究以深度學習之BERT技術提出一個工程訴訟案件篩選與歷審統計表建立及案件預測系統，並分為三個部分。第一部份是工程訴訟案件篩選，由中華民國司法院提供之判決書資料中經由基於BERT的模型架構篩選出屬於建築工程訴訟之案件，其準確率達到93.55%。第二部分是案件歷審統計表建立，將案件的歷審判決書利用正則表達式進行資訊擷取並彙整成個案之歷審統計表，準確率達到86.75%。第三部分是案件預測，利用基於多BERT的模型架構預測法院判決之結果，並找出相似的案例及同案件類型之統計表格，而判決預測在金額上及時間上準確率分別達到82.22%及88.89%。

This research uses the multiple BERT model to propose an construction litigation case screening and summary table creation and case prediction system, which is divided into three parts. The first part is the screening of construction litigation cases. From the judgment data provided by the Judicial Court of the Republic of China, the cases belonging to the construction litigation are selected through the BERT based model structure, and the accuracy rate reaches 93.55%. The second part is summary table creation, which uses regular expressions to extract information from the judgments and integrate them into a case summary table, with an accuracy rate of 86.75%. The third part is the case prediction. The multiple BERT model framework is used to predict the outcome of court judgments, and to find similar cases and statistical tables of the same case type. The accuracy rates of judgment prediction in terms of amount and time are respectively 82.22% and 88.89%.