Research in computational question answering has always been conducted with respect to specific tasks. Woods' LSNLIS program answers questions about its technical knowledge of moon rocks (see Woods 1972); Winograd's SHRDLU answers questions about the manipulation of blocks on a table (see Winograd 1972); Shortliffe's MYCIN helps physicians diagnose diseases and prescribe drugs by engaging in question answering dialogues (see Shortliffe 1974); Schank's SAM reads stories and answers questions about what it reads (see Schank et al. 1975). While all of these programs are concerned with question answering, each one has perceived the problem from a different angle.