Using Bidirectional Encoder Representations from Transformers (BERT) to predict criminal charges and sentences from...

Area of Science:

• Computational Linguistics
• Legal Informatics
• Artificial Intelligence

Background:

• Understanding legal statutes and predicting criminal behavior outcomes can be challenging for individuals unfamiliar with the law.
• Criminal justice systems generate vast amounts of textual data, such as court judgments, that hold valuable predictive information.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting criminal charges and sentence lengths using Taiwanese district court judgments.
• To adapt and improve the Bidirectional Encoder Representations from Transformers (BERT) model for analyzing legal texts and overcoming its inherent limitations, such as the 512-token limit.

Main Methods:

• Utilized a dataset of Taiwanese criminal judgments from district courts.
• Applied and fine-tuned the Bidirectional Encoder Representations from Transformers (BERT) model for two primary tasks: criminal charge prediction and sentence length prediction.
• Developed a novel solution to address BERT's 512-token input limitation for processing lengthy legal documents.

Main Results:

• Achieved a high accuracy of 98.95% in predicting criminal charges.
• Demonstrated feasibility of using BERT for analyzing Taiwanese criminal judgments.
• Obtained promising accuracy rates for sentence length prediction: 72.37% for injury trials and 80.93% for public endangerment trials.

Conclusions:

• The BERT model is a viable tool for analyzing legal texts and predicting outcomes in the Taiwanese criminal justice system.
• The proposed modifications effectively handle the limitations of BERT when applied to extensive legal documents.
• This research paves the way for AI-driven legal analytics and decision support systems.