Question answering in natural languages provides an intuitive and efficient way to help people access the rich information stored in various kinds of knowledge graphs (KGs). One of the key challenges for question answering over knowledge graphs (KGQA) is to learn a semantic representation of the input question and candidate relation chains over KGs and accurately measure the similarity between them. However, existing methods often failed to capture the semantic similarity for complex question answering, e.g., multi-hop and temporal constrained situations. In addition, existing KGQA related research mostly concentrates on entities while often ignores the events which contain a large portion of the world knowledge. To solve this issue, we propose a Contrastive Semantic Similarity Learning (CSSL) method for multi-hop question answering over event-centric KGs. In this method, for candidate relation chains generation, the retrieval subgraph is first constructed by identifying the topic event or entity in the question. To better accommodate complex questions, we introduce the contrastive learning framework to learn a common semantic space, where the similarity score is finally calculated to select the final answer. The experimental results on the EventQA dataset show that the proposed method achieves superior performances compared to the state-of-the-art baselines