The generation of referring expressions is one of the most extensively explored tasks in natural language generation, where a description that uniquely identifies an instance is to be provided. Some recent approaches aim to discover referring expressions in knowledge graphs. To limit the search space, existing approaches define quality measures based on the intuitiveness and simplicity of the discovered expressions. In this paper, we focus on referring expressions of interest for data linking task and present RE-miner, an algorithm tailored to automatically discover minimal and diverse referring expressions for all instances of a class in a knowledge graph. We experimentally demonstrate on several benchmark datasets that, compared to existing data linking tools, referring expressions for data linking substantially improve the results, especially the recall without decreasing the precision. We also show that the RE-miner algorithm can scale to datasets containing millions of facts.