Automação da análise na colidência visual de marcas: Uma abordagem com LLMs

Abstract

The visual similarity assessment of figurative trademarks is one of the most challenging tasks in trademark examination, as it requires the simultaneous comparison of multiple graphical elements and largely depends on the examiner's subjective interpretation. Given the growing number of trademark applications, this study aimed to develop and evaluate an automated method based on large multimodal language models to support the identification of visual similarity between figurative trademarks. The proposed methodology comprised four main stages: the construction of a derived dataset composed exclusively of pairs of figurative trademarks extracted from real Brazilian trademark cases; the fine-tuning of four open-source multimodal models using the Low-Rank Adaptation (LoRA) technique; the development of an explanation model guided by prompt engineering and few-shot learning; and the quantitative and qualitative evaluation of the results. Among the evaluated models, Qwen2 achieved the best performance, reaching an accuracy of 86% and demonstrating a high capability to identify actual trademark conflicts. Consequently, it was selected to generate technically grounded explanatory reports, which were positively evaluated by intellectual property specialists. The results indicate that the proposed method reduces subjectivity, enhances transparency, and represents a promising decision-support tool for figurative trademark similarity analysis, although human review remains essential.