This paper presents a sentence generation algorithm, which takes as input a context-free grammar and produces a set of sentences that achieves branch coverage for the grammar. The algorithm incorporates length control, redundancy elimination, and sentence-set size control strategies into a sentence generation process such that the generated sentences are short and simple, and the sentence set is small with no redundancy. The paper also investigates the application of this algorithm to test data generation for grammar-based systems. Experimental results show that the generated test data not only has high fault detection ability, but can also help testers improve the testing speed.