Explorando algoritmos de aprendizado de máquina em textos legais antitruste

Abstract

This research explores the integration of machine learning algorithms and NLP in antitrust analysis by CADE in Brazil, utilizing topic modeling to quantify how the prevalence of topics can assist in predicting decisions in cartel cases. The focus is on identifying the most efficient technique for examining legal texts from CADE, concentrating on understanding the decision-making process and evaluating relevant algorithms, including investigating hypotheses such as the potential superiority of the BERT model, particularly through BERTopic, in identifying topics in antitrust legal texts. The methodology encompasses the collection and analysis of procedural and biographical data from CADE authorities, employing various modeling tools like NMF, LDA, CTM, Top2Vec, and BERTopic, as well as metrics like NPMI, UMass Coherence, topic diversity, and processing time, taking ethical considerations into account. The results show that the BERTopic model, especially in BERTimbau and DistilUSE configurations, is notable in coherence, thematic diversity, and temporal efficiency, becoming a promising option for analyses in the context of CADE; the research emphasizes the importance of a careful selection of NLP models, ranging from LDA, ideal for high coherence and efficiency, to embedding-based models, more suitable for thematic diversity, and highlights limitations encountered, such as negative NPMI values, suggesting a need for improvement in topic coherence and precision of model settings. Moreover, the research explores the varied performance of different topic modeling techniques and the interplay between the sophistication of the techniques and the need for computational resources, highlighting the relevance of these approaches for the fields of Economics and Law and underscoring the value of applying advanced computational methods in these fields. In conclusion, the thesis emphasizes the importance of data preprocessing and the balance between NLP techniques and the availability of computational resources, confirming the effectiveness of BERTopic in topic modeling in legal contexts, despite the need for adjustments in coherence and configurations; and finally suggests the need for future investigations to enhance NLP and topic modeling techniques, aiming to expand their applicability and relevance.