Again we are adhering to the guidelines from the Framework Model (section Methods). The methods suggested are concerned with amount and composition of test material, and repeated here:
In the spirit of re-usability, we used test suites developed by the TSNLP project available on the www (LRE-62-089, Test Suites for Natural Language Processing). These test suites were constructed following the same guidelines, and all available examples relevant to the phenomena tested were retrieved.
Measuring precision and recall is done by running a grammar checker on the test material and counting the number of rejected and accepted items, respectively.