The optimum corpus sample size