Concepts in context: Evidence from a feature-norming study

Concepts in context: Evidence from a feature-norming study Diego Frassinelli (d.frassinelli@sms.ed.ac.uk) Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB, UK Alessandro Lenci (alessandro.lenci@ling.unipi.it) Dipartimento di Linguistica “T. Bolelli”, Via Santa Maria 36 56126 Pisa, Italy Abstract Concepts are typically conceived as context-free knowledge structures. Recently, a different view has emerged according to which subjects produce situation-specific conceptualizations, thereby raising important questions about the level of contex- tual dependency in conceptual representations. In this paper, we present a feature-norming study in which subjects are asked to generate properties of concepts presented in context. Col- lected data are analysed to investigate the actual amount of conceptual variation induced by contexts and the effect of con- text modality. Keywords: Semantic feature norms; property generation; con- text. Concepts and contexts Both in classical and in post-classical models, concepts have been conceived as substantially context-free knowledge struc- tures. Regardless of the particular theory (e.g. exemplar, pro- totype, and connectionist), it is generally assumed that con- cepts result from abstracting critical information about an en- tity per se (such as shape, colour, etc.), leaving behind back- ground situations (i.e. the contexts) in which these entities are experienced. Concepts thus become invariant to different contexts of use. Accordingly, the same representation of an apple is used both when categorizing an entity on a tree, and when categorizing the same entity in a supermarket. Recently, this view has been overtly criticized. For in- stance, Yeh and Barsalou (2006) argue that concepts not only contain a large array of situational information about the physical settings, events, and subjective perspectives of agents, but they also produce different conceptualizations in different contexts. For instance, the supermarket situa- tion would activate context-specific information concerning an apple, different from that activated by a different context, such a tree in a garden. These two claims directly follow from the perceptual simulation model adopted by the authors, but more in general they raise important questions about the level of contextual dependency in our conceptual represen- tations. Wu and Barsalou (2009) used a property genera- tion task to investigate the situated nature of concepts, and reported that approximately 26% of the features produced by subjects were indeed situation-related. Subjects gener- ated properties (semantic feature norms) provide interesting evidence about conceptual representations, but one intrinsic limit of the study in Wu and Barsalou (2009) is that stimuli were presented out of context, as it is customary in seman- tic norming. This way, it becomes impossible to address and test the more specific and crucial issue concerning the rela- tion between concepts and context, that is the actual effect of the context in modulating and biasing conceptual representa- tions. In this paper, we present a feature-norming study in which subjects are asked to generate properties of concepts pre- sented in context. To the best of our knowledge this is the first property generation task with this design. While we do not commit ourselves to any specific model of conceptual rep- resentation, collected data allow us to address directly three key issues concerning the effects of different contexts on con- cepts: i.) the actual amount of conceptual variation induced by contexts, and ii.) the property types that are more sub- ject to contextual variation, and iii.) the effect of the context modality. In particular, we will investigate the effect of both linguistic contexts (i.e. a sentence in which the context noun appears) and extralinguistic contexts (i.e. an image of a situ- ation in which an entity can be experienced). Semantic Feature Norms Nowadays there is a strong consensus on the fact that it is possible to describe the internal structure of a concept in terms of a set of semantic properties (Garrard, Lambon Ralph, Hodges, & Patterson, 2001; Baroni & Lenci, 2008). A tra- ditional way to access and study the structure of conceptual knowledge is the use of semantic features norms. These are lists of properties that participants produce describing and defining a specific concept; moreover they include several measures and statistics calculated according to feature pro- duction frequencies. As suggested by McRae and colleagues (McRae, Cree, Sei- denberg, & McNorgan, 2005) these lists do not provide a static and definitive representation of concepts, however, they are the most direct way to study the dynamics associated with the online process that takes place when subjects have to pro- cess a specific concept. Different researchers used these lists to investigate various aspects of human cognition. They have been used to test the psychological validity of cognitive theories (Wu & Barsalou, 2009), and as stimuli for different experiments such as seman- tic similarity (McRae, Sa, & Seidenberg, 1997) and property verification tasks (Cree, McNorgan, & McRae, 2006).