Distant supervision is a suitable method for relation extraction in big data. It provides a large amount of sample data by aligning relation instances in knowledge base with nature sentences in corpus. In this paper, a new method of distant supervision with expansion of ontology-based sampling is investigated to address the difficulty of extracting relations from sparse training data. First, an ontology which has a deep link with relation extraction is sought through the definition of cover ratio and volume ratio. Second, some relation instances are added by ontology reasoning and examples of queries. Finally, the expansion of training sets is completed by aligning the new relation instances and nature sentences in corpus. The experiment shows that the presented method is capable of extracting some relations whose training sets are weak, a task impossible by the normal distant supervision method.