A Formidable Ability: Detecting Adjectival Extremeness with DSMs

Abstract

While distributional semantic models (DSMs) can successfully capture the similarity structure within a semantic domain, less is known about their ability to represent abstract semantic properties that hold across domains. Such properties can form the basis for abstract semantic classes that are a crucial aspect of human semantic knowledge. For example, the abstract class of extreme adjectives (such as brilliant and freezing) spans a wide range of domains (here, INTELLIGENCE and TEMPERATURE). Using a model that compares query items to an aggregate DSM representation of a set of extreme adjectives, we show that novel adjectives can be classified accurately, supporting the insight that a cross-domain property like extremeness can be captured in a word's DSM representation. We then use the extremeness classifier to model the emergence of intensifier meaning in adverbs, demonstrating, in a separate task, the effectiveness of detecting this abstract semantic property.