Um método baseado em grandes modelos de linguagem generativos para a análise de similaridades entre marcas no contexto brasileiro

Abstract

A trademark aims to uniquely and distinctively identify the products and services offered by a company. It constitutes an important intangible asset, serving as a fundamental tool to prevent unfair competition practices and strengthen the company’s competitive position in the market. However, the increasing number of trademark registration applications submitted to the Brazilian National Institute of Industrial Property (INPI) has posed significant challenges, including longer review times, inconsistencies in decisions, and increased complexity in identifying trademark conflicts. In this context, automated methods for trademark similarity analysis have become essential to enhance the efficiency, reliability, and speed of the institute’s processes. This dissertation proposes a method based on Large Language Models (LLMs) to classify and explain the similarity between word marks, taking into account the phonetic, ideological, distinctive, and market-related criteria adopted by INPI. To achieve this goal, the method was structured into two main stages: (1) the development of a classification model capable of quickly and accurately identifying conflicts between trademarks; and (2) the development of an explanation model responsible for providing detailed justifications regarding why two trademarks are considered similar or not. To support this approach, a dataset named the Conflicting Marks Archive Dataset (CMAD) was created, comprising real trademark conflict cases extracted from INPI’s official publications. Several LLMs were evaluated for their accuracy in classifying conflicts and for generating well-founded explanations of the analyzed cases, which were validated by Intellectual Property (IP) specialists. The results showed that some of the evaluated models achieved outstanding performance, with accuracy close to 99%, F1-score above 98%, and AUC above 99% in the classification task. Furthermore, the explanations provided by the models received average ratings above 4 (on a scale from 0 to 5) from experts, demonstrating clarity and alignment with the criteria established by INPI. Therefore, this study contributes to the field of automated trademark similarity analysis by highlighting the potential of LLMs to optimize and enhance processes related to IP, reducing the subjectivity inherent to manual evaluation and increasing transparency in institutional decisions. The proposed method also facilitates broader access to trademark protection, especially for small and medium-sized enterprises facing financial or technical limitations,

thus helping reduce operational costs, prevent trademark disputes, and accelerate decision- making processes.