Analysing similarities between legal court documents using natural language processing approaches based on...

Area of Science:

• Computational Linguistics
• Artificial Intelligence in Law
• Natural Language Processing

Background:

• Judicial systems face challenges with large volumes of legal documents.
• Natural Language Processing (NLP) offers potential solutions for managing legal data.
• Transformer-based AI models have shown promise in various NLP tasks.

Purpose of the Study:

• To evaluate transformer-based NLP models for detecting similarity among judicial documents.
• To compare the performance of different NLP techniques in the Brazilian legal context.
• To investigate the application of AI in expediting judicial proceedings.

Main Methods:

• Employed eight NLP techniques, focusing on transformer architectures (BERT, GPT-2, RoBERTa, LlaMA).
• Pre-trained models on Brazilian Portuguese corpora and fine-tuned them on 210,000 legal cases.
• Generated document vector representations using embeddings and evaluated similarity via cosine distance.

Main Results:

• Transformer-based models significantly outperformed traditional NLP techniques.
• The LlaMA model, fine-tuned for the legal domain, achieved the highest accuracy in document similarity detection.
• Demonstrated a robust methodology for analyzing large legal document datasets.

Conclusions:

• Transformer models are highly effective for NLP tasks in the legal sector, particularly for document similarity.
• Fine-tuning models like LlaMA for specific domains, such as Brazilian law, enhances performance.
• This research advances AI applications in law and supports Sustainable Development Goals.