Combined classes of three slots
In addition, we compared the models with the three slots taken together. The 5 to 100 classes of the Effected Predicates were combined with 2, 5,10,15, 20 and 30 classes of the Causer and the Causee. The best-performing models only were tested: depRel8 for the Causer and 23syn for the Effected Predicate. For the Causee, also depRel8 was chosen, although it did not perform better than the bag-of-words model. The C index for these different classifications is displayed in Figure 5.
The results of the combined classifications show that adding the semantic information about the Causer and Causee to the information about the Effected Predicate does increase the predictive power, but the effect is non-additive. The effect is significant for the small number of the Effected Predicates, but as the verb classes become finer-grained and more successful in the prediction, the impact of the nominal classes becomes smaller. Also the difference in the granularity of the nominal slot classifications becomes less evident as the number of the verb classes and the predictive power grow.
Figure 5: Thee slots together: predictive power of the best-performing individual classifications
-  Unfortunately, we were unable to test the statistical interactions of the slot fillers due to datasparseness.