Comparison of Feature Extraction Methods for Brazilian Legal Documents Clustering

This article aims to evaluate the impact of different textual feature extraction methods in the task of clustering Brazilian legal texts. We compared Binary Bag of Words, Bag of Words, Term Frequency-Inverse Document Frequency, Word2vec and Doc2vec models in different dimensions and with different hyperparameters, totaling 45 models. Our experiment consists in evaluating the result of clustering done by K-Means algorithm over the vectors created by each model. The evaluation was done both quantitatively, using clustering evaluation metrics, and qualitatively, considering relevant aspects for the application of this type of algorithm in the legal environment, such as transparency and interpretability. Our experiments were conducted in a database of 30,000 documents in Brazilian Portuguese of judicial moves of the Tribunal de Justiça do Rio Grande do Norte (TJRN). The research results suggest that the TF-IDF method seems to be the most suitable for the task, outperforming the other models in considered metrics. The other methods appear to perform equally well, with the exception of Doc2vec, which performed poorly.