This project uses opportunistic abilities in two ways. First, the DET preconditions allow the failing of a rule based on the opportunistic gathering of information related to the semantic structure of the input. In the vacation example, this might include something like calling ahead to see if Aunt Millie is going to be home on her birthday, rather than just showing up there and finding an empty house. In RELATION-20 above, one of the DET preconditions in rule 0 is that the head proposition's agent was the same as the rest (supporting) proposition's agent. This is a simple example, but it can be extended to include any kind of reference to the input semantic structure or the possible realizations available. Lexical availability of a verb form such as a nominalized form, handled by the DEP rules in the example, could also be determined using the DET rules.
The advantage to this opportunistic use of information gathering using DET rules is
that it prevents the needless application of impossible rules that would eventually
lead to backtracking. Instead of assuming that the two propositions' agents were
equal, planning the whole left side of the tree in Figure 1, and then having to
backtrack when the proposition is reached on the right side of the tree
, the DET rule will automatically fail the rule before the
recursive descent algorithm even begins. This prunes the input search space, as well
as eliminates needless backtracking.
The second way this system is opportunistic is a little more subtle. It isn't really ``opportunistic'' in the sense that it performs special processing based on a current need (such as described above for the DET rules); however, it produces what are generally considered to be opportunistic effects. The system can use DEP rules and constraints to eliminate conflicting sets of rules. The elimination of RELATION-20 rule 0 because no progressive surface form was available in the DEATH-1 proposition is an example. DEP preconditions can be written in such a way as to specify the constraints necessary for the rule to be applied. Because the rule set is instantiated based on the input semantics, different rules will be present depending on what the input is. Because different rules are available, different sets of constraints will also be available. These constraints can be used to signal the availability of different forms or the presence of any other condition which might be helpful in determining the applicability of a rule, and the DEP rules then can act as ``demons'' that either reach out and pick up their needed constraints when available, or cause the rule to fail if the constraints are not available.
Because the rule set dependencies are pre-analyzed, no special action is required; all rules whose DEP ``demons'' cannot find their constraints are failed automatically. However, this can be seen simply as an optimization of opportunistic planning. Instead of having to leave the current process to find information or perform other processing, the results are returned automatically.
This second type of opportunistic ability is helpful even during the modified recursive
descent. As the recursive descent process makes choices, the problem becomes more and
more constrained as rules are eliminated. As
rules are eliminated, some of the DEP demons may find the constraints they want are no longer
available. This situation is automatically detected and any such rules are also failed.
Thus, this system has a very sophisticated opportunistic ability which continually updates
the opportunistic processes. Not only is something like the availability of a surface
progressive form checked at the beginning of processing, but if such a form becomes
unavailable due to any reason during processing, it will be automatically detected.