Generalized Zero-Shot Learning (GZSL) in computer vision helps improve swarm intelligence decision-making with limited labeled data. Synthesizing pseudo samples for unseen classes in GZSL is effective, but recent methods ignore the role of learned task-irrelevant features when synthesizing. These features have better transferability and can improve the generalization of models. In this paper, we propose a Task-irrelevant Context Enhanced Pseudo Sample Synthesis (TCEPS \(^2\) ) with two modules, which synthesizes pseudo samples by taking advantage of task-relevant features and task-irrelevant contexts simultaneously. And the task-irrelevant contexts consist of continuous task-irrelevant features and discrete task-irrelevant concepts. In the Task-irrelevant Contexts Extraction Module, we disentangle the global visual features into task-relevant features, task-irrelevant features and task-irrelevant concepts. In the Pseudo Sample Synthesis Module, we aim to synthesize different types of pseudo samples for conformity, diversity, and rationality. Extensive experiments show our methods are competitive on different benchmarks and tasks.