Increasing diversity: Natural language measures for software fault prediction