Refining WordNet adjective dumbbells using intensity relations

We propose a new semantic relation of intensity for gradable adjectives in WordNet and show its specific benefits for NLP. Intensity enriches the present, vague, similar relation with information on the degree with which different adjectives express a shared attribute. Using lexical-semantic patterns, we mine the Web for evidence of the relative strength of adjectives like “large”, “huge” and “gigantic” with respect to their attribute (“size”). The pairwise orderings we derive allow us to construct scales on which the adjectives are located. To represent the intensity relation among gradable adjectives in WordNet, we combine ordered scales with the current WordNet dumbbells based on the relation between a pair of central adjectives and a group of undifferentiated semantically similar adjectives. A new intensity relation links the adjectives in the dumbbells and their concurrent representation on the scale.