A Crow's Beak is not Yellow: Investigations on Cognitively Salient Concept Properties

Subject-generated concept descriptions in terms of properties of different kinds (category: rabbits are mammals, parts: they have long ears, behaviour: they jump, ...) are widely used in cognitive science as proxies to feature-based representations of concepts in the mind. These feature norms (as collections of subject-elicited properties are called in the relevant literature) are used in simulations of cognitive tasks and experimental design. Moreover, vector spaces that have subject-generated properties as dimensions have been shown to be a good complement or alternative to traditional semantic models based on corpus collocates. Since the concept–property pairs in feature norms resemble the tuples that semantic relation extraction algorithms extract from corpora, recent research has attempted to extract feature-norm-like concept descriptions from corpora. From a practical point of view, the success of this enterprise would mean being able to produce much larger norms without the need to resort to expensive and time-consuming elicitation experiments, leading to wider cognitive simulations and possibly better vector space models of semantics. Lexical resources incorporating semantic relations between lexical entries would profit likewise from such automatic extraction methods that would facilitate extending the lexical resource with relation instances that are prominent to speakers from a cognitive perspective. From a theoretical point of view, a corpus-based system that produces human-like concept descriptions might provide cues of how humans themselves come up with such descriptions. The general goals of this dissertation are (i) to report empirical investigations of the cognitive salience of semantic relation types and (ii) to present a case study about extracting cognitively salient concept properties from text corpora, namely composite expressions for constitutive parts of concepts (e. g., crow: has a black beak).