Relation classification (RC) is commonly the second step in a relation extraction pipeline, which asserts the relation of two identified entities based on their context. The latest trend for dealing with the task resorts to pre-trained language models (PLMs). It transforms the discriminative RC into a linguistics problem and fully induces the language knowledge PLMs derived from pre-training. Despite the visible progress, existing approaches handle only one relation between each entity pair while workless in real cases where multiple relations may be valid, i.e., entity pair overlap (EPO), leading to their limited applications. In this paper, we introduce ConFit, a novel contrastive learning based approach that fine-tunes text-to-text PLM for relation classification. ConFit reformulates RC as a restoration problem of textualized relations, and it learns to bind sequential words of a candidate relation with a probability mass above or below a threshold, corresponding to whether the relation truly holds. As a result, the learned model adaptively phrases diverse relations through a decoding, scoring, and selection workflow to fit EPO scenarios. Extensive experiments on four widely used datasets evidence that T5-large fine-tuned with ConFit significantly outperforms previous methods, whether single or multiple relations exist.