Distinguishing Properties and Relations in the Denotation of Adjectives: An Empirical Investigation

Abstract

We empirically investigate the task of classifying adjectives into property-denoting vs. relational types, a distinction that is highly relevant for ontology learning. The feasibility of this task is evaluated in two experiments: (i) a corpus study based on human annotations and (ii) an automatic classification experiment. We observe that token-level annotation of these classes is expensive and difficult. Yet, a careful corpus analysis reveals that adjective classes tend to be stable on the type level, with few occurrences of class shifts observed at the token level. As a consequence, we opt for an automatic classification approach that operates on the type level. Training on heuristically labeled data yields high classification performance on our own data and on a data set compiled from WordNet. Our results indicate that it is feasible to automatically distinguish property-denoting and relational adjectives, even if only small amounts of annotated data are available. A combination of semantic, morphological and shallow syntactic features turns out to be most informative for the task.