Template Filling

Once you're done parsing, you should have basic semantic units.
Coreference resolution may have decorated these, so that you have recorded information about items.
You then need to, in a task specific way, fill in your templates.
In essence, this is where you query the knowledge you've got from NLP to build your answers.
If my templates are about who eats chocolate, I can look for sentences like "Chris eats choocolate."
Depending on the parser, these may be recorded semantically or syntactically.
The system can also bridge items by guessing.
For example, if you've not got a full parse of a sentence, but two items are in the same sentence, you might want to guess that their relationship is enough to fill in a template.
You can imagine how this is easier for just named entity categorisation.