Comparing CNNs and Transformers for Multiclass Patent Classification in Spanish

Automating patent classification in the International Patent Classification system is a challenging task due to the complexity and diversity of Spanish patent texts. This study compares the performance of Convolutional Neural Networks and Transformers, leveraging BETO, a Spanish-specific BERT variant, on a large dataset from PATENTSCOPE. By evaluating accuracy, processing time, and computational efficiency, we provide insights into the strengths and limitations of each approach, offering recommendations for text classification under hardware constraints. Our results reveal that Transformers outperform CNNs in classification accuracy, particularly for highly diverse text inputs, due to their ability to model complex contextual relationships. However, CNNs demonstrate faster processing times and lower computational demands, making them more suitable for environments with limited hardware resources. These findings underscore the importance of balancing model performance with resource constraints when selecting architectures for real-world applications. These findings aim to enhance the efficiency of automated patent management in Spanish, facilitating better handling of technological information in multilingual contexts.