Assessing rule-based document segmentation and word normalization for legal ruling classification

The Brazilian judiciary has shown promising interest in applying Natural Language Processing (NLP) techniques to various legal tasks. One such application is classifying legal rulings by the topics of recurring appeals. This study investigates two key strategies for preprocessing legal documents, drawing on insights from legal domain experts: whether using specific sections of the document is more effective for legal ruling classification than analyzing the entire document, and which expressions can be normalized to standardize the document vocabulary. The experimental results indicate that combining normalization preprocessing with the extraction of the judge’s manifestation section yields better performance, as measured by the F1 score. Additionally, we demonstrate how the Jaccard similarity index provides valuable insight into the impact of the preprocessing pipeline on the TF-IDF feature extraction method and, by extension, on document representation. This paper underscores the importance of leveraging domain expertise to guide an optimal set of preprocessing operations.