A.5.1 Chapter 3.3 Charts

The following three subsections present the error summary files as generated by the restore.input program (see Appendix 4). The files show the performance of the network, for tolerance values of 0 to 0.5, on each of the 17 output units. The format for each unit is the following:

*** Name of unit (ABS: absolute number of times the unit should be 1 in the corpus)

• On: Absolute number of times the unit should have been on, but was not.
• Mis: The percentage of missed errors, as calculated by dividing the value for On by ABS -- rounded down).
• Off: Absolute number of times the unit should have been off, but was not.
• Spu: The percentage of spurious errors, as calculated by dividing the value for Off by the absolute number of times the unit should be 0 -- rounded down. The latter absolute number can be found by subtracting ABS from the total number of patterns (this number is printed near the end of the file).
• Bad: The percentage found by first adding On and Off and then dividing this sum by ABS. (These values are never used in 3.3 or 3.4).
• Gud: The percentage found by first adding On and Off and then dividing this sum by the total number of patterns. (These values are used in 3.3).

The CST percentages are the averages for the 7 clause types.

At the end of each summary file, a table gives the absolute number of the input patterns which occur in clauses of varying depth (depth 0 patterns are punctuation marks), and the absolute numbers and percentages of pattens which are correctly and incorrectly classified for these depths, as calculated by checking units ID1, ID2 and ID3 simultaneously for a tolerance value of 0.2 for each of the units.