Transformer-based embeddings and stacking classifiers for judgment outcome prediction in multilingual legal texts

Area of Science:

• Computational Linguistics
• Legal Informatics
• Artificial Intelligence

Background:

• Commonwealth legal systems predominantly use English, creating a language barrier for non-English speakers.
• This linguistic divide hinders access to justice, limiting understanding and utilization of legal services.
• Existing legal outcome prediction models often lack multilingual capabilities.

Purpose of the Study:

• To develop a novel approach for predicting legal outcomes using summarized legal texts in English and Indian regional languages (Kannada, Tamil, Telugu).
• To bridge the linguistic divide within the Indian judicial system.
• To evaluate the effectiveness of multilingual summaries in legal outcome prediction (LOP).

Main Methods:

• Utilized an embedding module with RoBERTa, FastText, InLegalBERT, and IndicBERT.
• Employed a multi-stage classification module with classifiers including Random Forest, Decision Trees, SVM, KNN, MLP, XGBoost, LightGBM, and Naive Bayes.
• Incorporated metamodels to mitigate overfitting and underfitting issues.

Main Results:

• Achieved a test accuracy of 91.45% and a train accuracy of 94.10%.
• Obtained F1 scores of 0.91 for testing and 0.94 for training.
• Demonstrated that summarized texts in regional languages are effective for legal outcome prediction.

Conclusions:

• Introduced a summary-based, multilingual legal outcome prediction framework.
• Showcased the effectiveness of stacked ensemble learning in enhancing predictive performance and robustness.
• Promoted inclusivity and trust in the legal system for regional language speakers, advancing legal analytics and access to justice.