Understanding the lexical relationships between word pairs is paramount for advancing applications in natural language processing. However, existing methodologies often oversimplify these relationships, treating them merely as binary classifications. This approach overlooks the nuanced degrees of lexical relatedness that exist on a continuous spectrum. In this paper, we propose a novel lexical relation embedding approach grounded in hierarchical contrastive learning. By integrating instance contrastive learning and prototypical contrastive learning, our model effectively captures the hierarchical structure inherent in lexical relationships. Moreover, we introduce a ranking loss function to explicitly align the representation degrees of word pairs along the continuous dimension. Experimental results across three lexical-related tasks demonstrate the efficacy of our approach, achieving state-of-the-art performance on most datasets.