Object Activation from Features in the Semantic System
Michael A. Kraut, Sarah Kremen, Jessica B. Segal,
Vincent Calhoun, Lauren R. Moo, and John Hart, Jr.
Abstract
& The human brain is thought to elicit an object representa-
tion via co-activation of neural regions that encode various
object features. The cortical regions and mechanisms involved
in this process have never been elucidated for the semantic
system. We used functional magnetic resonance imaging
(fMRI) to evaluate regions activated during a task designed
to elicit object activation within the semantic system (e.g.,
presenting the words ‘‘desert’’ and ‘‘humps’’ with the task to
determine if they combine to form an object, in this case a
‘‘camel’’). There were signal changes in the thalamus for word
pairs that activated an object, but not for pairs that (a) failed to
activate an object, (b) were simply semantically associated, or
(c) were members of the same category. These results suggest
that the thalamus has a critical role in coordinating the cortical
activity required for activating an object concept in the
semantic system. &
INTRODUCTION
Numerous models of semantic memory (Millikan, 1998;
Schyns, Goldstone, & Thibaut, 1998; Lambert &
Shanks, 1997; Estes, 1994; Hinton & Anderson, 1989;
Fodor & Pylyshyn, 1988; Paivio, 1986; Collins & Loftus,
1975; Smith, Shoben, & Rips, 1974) have suggested the
existence of specific subprocesses, including, but not
limited to, making semantic choices from among multi-
ple possibilities, feature- object correlation, noun- verb
association, and adjective- noun association (Ricci et al.,
1999; Thompson-Schill, D’Esposito, Aguirre, & Farah,
1997; Demb et al., 1995; Kapur et al., 1994; Saffran &
Schwartz, 1994; Demonet et al., 1992; Petersen, Fox,
Posner, Mintun, & Raichle, 1988; Petersen, Fox, Snyder,
& Raichle, 1990; Posner, Petersen, Fox, & Raichle, 1988;
Shallice, 1988; Saffran, 1982). One other related proc-
ess that has been investigated via a variety of tasks is
semantic association. In a typical semantic association
paradigm, two items are presented with the general
instructions/question of, ‘‘Are these two stimuli related/
associated with each other?’’ The terms ‘‘related’’ or
‘‘associated’’ do not typically require a formal definition
or instruction to the subject; their meaning or intent is
conveyed by an example. (‘‘For instance, salt and
pepper are related.’’)
Within one theoretical framework, semantic associa-
tions have been shown to be behaviorally and anatom-
ically dissociable into two major subtypes: (1)
compositional associations (two items are associated
but retain their individuality in the relationship—salt
and pepper), and (2) noncompositional associations
(the two items become fused semantically—‘‘computer’’
and ‘‘virus’’ forming ‘‘computer virus’’) (Kounios, Smith,
Yang, Bachman, & D’Esposito, 2001; Fodor & Pylyshyn,
1988). Of interest, in the compositional associations, the
two related items remain distinct both lexically and
semantically, while in the noncompositional associa-
tions, the two stimuli fuse to form a third novel stimulus
(as in the ‘‘computer virus’’ example) at the semantic,
but not at the lexical, level. This concept of semantic
fusion, where two entities combine to form a new
representation, can be applied to other semantic rela-
tionships. One such application has been the proposal
that features that are common to an object could ‘‘fuse’’
to activate a previously stored semantic object represen-
tation (Gray, 1999; Eichenbaum & Bunsey, 1995; Singer
& Gray, 1995; Damasio, 1989, 1990; Warrington & Mc-
Carthy, 1987; Allport, 1985). While this semantic fusion
process has been proposed in theoretical form, a para-
digm addressing it during object activation in semantic
memory has not been formally developed.
To address the question of object activation within
this framework, we have designed an analog of the
semantic fusion task (a noncompositional association)
where the stimuli are two features that are common to
the target object. The particular features were chosen
such that, in combination, they are both sufficient and
specific to elicit the target object in semantic memory.
The task instructions are to push a button if the two
features (e.g., ‘‘desert’’ and ‘‘humps’’) combine to make
you think of an object (e.g., ‘‘camel’’). These task
instructions are in contrast, yet analogous to, those of
other semantic association tasks (‘‘Are these items re- Johns Hopkins University
© 2002 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 14:1, pp. 24- 36