Number of sentences and training noun compound instances after agreement selected from each text collection considered

The inter-annotator agreement on the semantic annotation of the noun compounds in the unshuffled training corpora

For the noun compound instances that encoded more than one semantic classification category, the agreement was done on one of the relations only. The agreement on the semantic relations for three noun compounds was computed on gold bracketing. “NA” means not available (due to time constraints, we were unable to annotate the set of three noun compounds with prepositional paraphrases).

The distribution of semantic relations on the annotated unshuffled training corpora after agreement

The semantic relations for which there was no example given were not encoded by the noun compounds.

Mapping between the two sets of semantic classification categories: 8 prepositional paraphrases (PPs) and 35 semantic relations (SRs)

The mapping was obtained on the unshuffled training corpora.

The performance obtained by the supervised models on Lauer’s test data (unshuffled) and, respectively, on the random test data (shuffled) for the interpretation task

Two sets of semantic classification categories were considered: 8 PPs (8 prepositional paraphrases), and 35 SRs (35 semantic relations). ‘SVM (+PP)” employs the same feature set as the SVM model plus the corresponding prepositional paraphrase.

Experimental results with Lapata and Keller’s web-based unsupervised interpretation model on different types of test sets

“No” means not ambiguous and “Yes” means ambiguous.

The performance obtained by the supervised and unsupervised probabilistic models on the two test data sets for the interpretation task

“C5.0 (+SR#i)” means the supervised model applied to the basic feature set plus SR#i as feature.