Estimating Upper and Lower Bounds on the Performance of Word-Sense Disambiguation Programs