Automatic similarity detection for trademark

Trademark opposition occurs when a newly filed trademark application is challenged for potential conflict with earlier rights, requiring a detailed comparison of texts describing protected goods or services. Using a dataset of 107,570 EUIPO records (2012–2024) describing office practices and decisions on pairs of goods and services, this paper presents different models to automatically predict the legal decision on similarity. The first contribution of our approach is to fine-tune embeddings from transformers and use them to vectorize the textual description of protected goods and services. The second contribution concerns a proposal of a hybrid neural network model that incorporates both the model embedding and categorical variables describing the context in which the legal decision was made. The goal is to provide a model able to learn both the semantics of the text description and the institutional rules for determining the similarity of the protected goods and services. We show that our model significantly improves on the state of the art, with more than 90% of F1-score, where previous approaches do not exceed 70%. We also propose lighter more explainable models (Decision Trees, Logistic Regression, Random Forest), with F1-score between 80% and 90%, that can be used by experts to justify justice decisions. • Fine tuning of Large Language Model embeddings for protected goods and services. • Hybrid model with embeddings and contextual categorical features. • Comparison and training with a dataset of 107,570 practices and decisions from EUIPO.