Análise de colisão de marcas nominativas: Um estudo sobre o desempenho de modelos de Processamento de Linguagem Natural (PLN)

Abstract

This study analyzed the effectiveness of different Natural Language Processing (NLP) models in assessing the risk of collisions between word marks. Three common NLP models for text classification were analyzed - Word2Vec, BERT and FastText - and applied to 304 word mark cases from the National Institute of Industrial Property (INPI) database that were denied registration due to collisions. The models' Accuracy, Precision, Recall and F1-score metrics were calculated to assess their performance in identifying collision risks. The results showed that although all the models achieved 100% Recall when flagging all collision cases, BERT performed best with 86.51% Accuracy and an F1-score of 92.77%, correctly classifying the majority of cases. Word2Vec showed lower Accuracy and Precision compared to the other models. FastText also provided relatively accurate results. This study advances understanding of the application of NLP approaches such as Word2Vec, BERT and FastText to semantically analyze trademark similarity. The findings offer practical guidance to practitioners on how to check collisions efficiently and quickly using NLP techniques.