Most of the existing studies on event coreference resolution utilize the trigger as the core semantic representation of an event. However, the presence of abstract expressions, such as interrogative words and quantifiers, within the annotation information of triggers, poses a challenge to the language models in comprehending the core semantics of events. Moreover, previous studies mostly focus on the training of identifying coreference relations, neglecting to learn the distinction between coreference and non-coreference. To address these two issues, we propose a document-level event coreference resolution method based on trigger augmentation and contrastive learning. Large Language Models (LLMs) are employed to revise the event triggers, augmenting their semantical information. Furthermore, we introduce the contrastive learning based on the relationships embedding for the first time, instructing the language model to distinguish between the coreference and non-coreference. Experimental results on the KBP corpus demonstrate that our proposed method outperforms the state-of-the-art baselines under both the annotated and the predicted events settings.