Aplicação e comparação de técnicas de classificação automática de documentos: um estudo de caso com o dataset do domínio jurídico “Victor”

Abstract

The application of Natural Language Processing (NLP) and Artificial Intelligence (AI) in the Brazilian legal context is a rapidly growing area that can alter the way legal professionals work, given the volume of generated text. Among the possible applications of NLP and AI is the automatic classification of documents, which, among other things, can be employed in the automation of the digitization process of Judicial Proceedings that are still in physical form. Therefore, this work applies and compares AI algorithms for the classification of legal documents. The algorithms are divided into two different approaches. The first approach (I) separates the computational representation process of the text from the classifier training itself and applies SVM and Logistic Regression in conjunction with computational representations based on TF-IDF, Word2Vec, FastText, and BERT. The second approach (II) simultaneously performs the computational representation of documents and the training of the classifier, applying Deep Learning algorithms based on recurrent neural networks, specifically ULMFiT (Universal Language Model Fine-tuning), and HAN (Hierarchical Attention Networks). The studied dataset is named VICTOR, composed of documents from the Supreme Federal Court (STF) of Brazil. The research concludes that both approaches can be applied to the classification of legal documents from the employed dataset. Additionally, despite being less computationally expensive, the classification pipelines of Approach I, which use the computational representation of the document with TF-IDF, yield results equivalent to pipelines employing Deep Learning. Furthermore, embedding documents specialization with data from the dataset under study, improves the performance of pipelines that employ Word2Vec, FastText and ULMFiT, compared to pipelines that apply the generic representations of these, i.e., models pre-trained with data from the general context.