Academic journal article Psychonomic Bulletin & Review

Evaluating the Random Representation Assumption of Lexical Semantics in Cognitive Models

Academic journal article Psychonomic Bulletin & Review

Evaluating the Random Representation Assumption of Lexical Semantics in Cognitive Models

Article excerpt

A common assumption implicit in cognitive models is that lexical semantics can be approximated by using randomly generated representations to stand in for word meaning. However, the use of random representations contains the hidden assumption that semantic similarity is symmetrically distributed across randomly selected words or between instances within a semantic category. We evaluated this assumption by computing similarity distributions for randomly selected words from a number of well-known semantic measures and comparing them with the distributions from random representations commonly used in cognitive models. The similarity distributions from all semantic measures were positively skewed compared with the symmetric normal distributions assumed by random representations. We discuss potential consequences that this false assumption may have for conclusions drawn from process models that use random representations.

A model of a cognitive phenomenon typically requires accounts of representation, of process, and of how the two interact (Estes, 1975). These two aspects of a model are interdependent, with the process requiring a representation on which to operate and the representation requiring a process to simulate behavior. For example, Rumelhart and McClelland (1982) created a model in which 16-feature vectors were used to represent capital letters in which each feature was the presence or absence of a line at a particular orientation, and they evaluated the process of interactive activation. The use of this representation was justified by research on how the visual system responds to primitive features. Similarly, some models use realistic representations of faces (e.g., Dailey & Cottrell, 1999) and orthographic and phonological characteristics of words (e.g., Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) or digits (e.g., Hinton, 2007).

If insufficient research exists to point to the correct representation, a common practice in cognitive modeling is to use randomly generated representations to stand in for psychological structure. This practice makes it unlikely that the representation is biased toward supporting the process mechanism, and the model can be refined later when further research reveals the correct representation. An example is Hintzman's (1986) use of random representations to simulate schema abstraction using Posner and Keele's (1968) stimuli: Stimuli were random dot patterns, and exemplars of the same category were random perturbations of a prototype pattern. Hintzman (1986) was able to create equivalent structure in his simulation by generating prototypes as random vectors and generating exemplars within a category as distortions of a prototype.

However, caution is needed if random representations are used. The performance of cognitive models is dependent largely on valid representational assumptions. For example, Daugherty and Seidenberg's (1992) connectionist model was only able to correctly simulate past tense verb processing when the model was trained on representations that contained the correct distributional structure. Alternatively, it is also possible for a process model to give a good account of the human behavior only when random representations are used, but not when the correct representational structure is encoded. Cree, McRae, and McNorgan (1999) argued that using a plausible representational structure, rather than random representations, constrains the modeling exercise by reducing degrees of freedom.

Random representations are commonly used in models of episodic memory. In global matching models of recognition memory (e.g., Hintzman, 1988; Murdock, 1982; Shiffrin & Steyvers, 1997), decisions are made by assessing the similarity of the probe word to the (usually noisy) study items with particular processing mechanisms. The use of random representations in these models produces a hidden assumption that the distribution of similarity across randomly selected words is symmetric and is approximately Gaussian. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.