Future Research

Almost as exciting as the project as it stands are the possible modifications that can improve it. These modifications are summarized below:

1. Allow for more than one set of rules per ss-node. At present, only one set of rules is allowed per input node. It would be extremely helpful to allow more than one set of rules, from each of which a rule would be (optionally) chosen. Each of these rule sets could interact with the other, with one mandatory set producing the output structure, and the others adding optional features. These rule sets could be opportunistically inserted depending on semantic input. The following benefits would be gained:
   • simplify rules
   • allows different tests to be applied depending on input, for instance

     planning attitudes

     planning focused elements

     planning modifiers

2. Implement dependency-directed backtracking. Instead of simply backtracking to the last node, we could backtrack to the cause of failure. All of the dependency information is already calculated, so this would be a relatively easy, though extremely useful, addition.

3. Keep track of surface ordering decisions so that sentence boundary decisions can be made more intelligently. See the discussion on ``completeness'' in the previous section.

4. Implement heuristics for choosing which rule to instantiate in modified recursive descent (i.e. pick most constraining rule). Instead of simply picking the first rule, pick the rule most likely to result in the best plan. This is one of the most general AI methods for improving efficiency, and it should definitely be implemented here (or at the very least, the system should allow for heuristics to be inserted).

5. Identify ``circuits'' of inter-dependency and create macro operators to lessen combinatorial explosion. In an application such as text planning, most of the constraints are local to a sub-tree of the search space. Thus, during backtracking, the optimal plan for a set of nodes may be calculated again and again. Instead of this, it would be helpful to identify these clusters of inter-dependent nodes, determine their optimal configuration off-line, and then during backtracking, they can be treated as a macro node.

6. Scale-up lexicon and rule-set. To better evaluate the system, the rule set must be expanded and more general semantic inputs used.

7. Apply method to other planning domains, such as content planning. It will be very interesting to apply these same techniques to other planning domains. In general, any problem that has inter-acting constraints would benefit from this approach.