Natural Language Processing mainly relies on large language models requiring very large datasets and behaves like a “black box”. These models are the baselines for many classification tasks, such as disinformation detection. Recently, hybrid approaches between deep learning and symbolic AI try to outperform attention-based models by introducing symbolic reasoning in the decision process, making it more understandable to the users. In this paper, we introduce CATS, an explainable attention mechanism based on the semantic analysis of documents. This approach outperforms equivalent fully-neuronal models, reduces annotated data needs and allows to understand how the decision process is made.