Using instruments to understand argument structure: Evidence for gradient representation