A trademark protects a sign for goods and services, which are described in words or groups of words called items. When a trademark owner identifies a mark similar to his registered trademark, he can file an opposition. Since the goods and services of each trademark are described by hundreds of items, opposition requires a complex manual analysis of the similarity between the goods and servicesof the two trademarks. The aim is to propose an automatic method for finding similarities between trademarks using LLM embeddings on items. The cosine similarity measure is then integrated with other features in a supervised learning process to predict similarity. A history of over 120,000 European Trademark Office decisions is used to assess the quality of the results.