Identificação automática de medicamentos em textos da justiça federal do Rio Grande do Norte com base em técnicas de processamento de linguagem natural

Abstract

This work aims to automatically identify and quantify requests for medications in legal cases from the Federal Small Claims Court (JEFs) in Rio Grande do Norte, Brazil, using Natural Language Processing (NLP) and Machine Learning (ML) techniques. We extracted 65,875 documents from the JEFs' CRETA system and refined them to a balanced subset between initial petitions and sentences (n=11,364). We performed data cleaning, treatment and label review. We created a weak supervision pipeline to label records regarding involvement with healthcare (n=6,196). We evaluated BioBERTpt models for named entity recognition to detect medications. We applied the best model to extract terms and associated them with ANVISA/CMED tables for standardization and pricing. Finally, we developed Qlik Sense dashboards to quantify expenses and visualize medication demands. The main challenges were computational constraints and difficulties in model evaluation due to problems with the manually labeled sample. We conclude that NLP and ML techniques have great potential to extract insights from legal cases.