In this study, the importance of patch representation in automating various software engineering tasks is emphasized. Existing approaches utilizing deep learning to represent patches focus on token sequences or Abstract Syntax Trees (ASTs) but lack explicit calculation of intention semantics induced by changes on the AST. The proposed methodology, referred to as Patcherizer, introduces an innovative approach that incorporates intention features of context and structure, capturing sequence intention and graph intention within code changes. Patcherizer utilizes graph convolutional neural networks for structural input representation and transformers for intention sequence representation, achieving superior performance across tasks such as generating patch descriptions in natural language. Experimental results demonstrate Patcherizer’s effectiveness, with significant improvements over state-of-the-art approaches on the task of patch description generation.