Semantic Memory 451for these phenomena were necessarily more focused than
were the original models of semantic memory. In this section
of the chapter, we take a quick look at two more recent ap-
proaches to understanding knowledge representations.
Distributed Network Models
Distributed network models have a long history (e.g., Hebb,
1949; Rosenblatt, 1962), but they did not become influential in
cognitive psychology until the mid-1980s (e.g., McClelland &
Rumelhart, 1986; Rumelhart & McClelland, 1986). The de-
velopment and investigation of distributed network models
has become a gigantic enterprise. Our goal will be to summa-
rize the most important characteristics of these models,
especially as they apply to semantic memory. According to
distributed network models, concepts are represented as pat-
terns of activation across a network of densely interconnected
units. Similar concepts are represented by similar patterns of
activation. The units can be thought of as representing aspects
of the object or event being represented. These aspects, how-
ever, need not be nameable or correspond in any obvious way
to the features people might list in a description of the entity.
Indeed, a traditional feature, such ashas wings,might itself be
a pattern of activation over a collection of units.
Units are typically organized into modules, which corre-
spond to sets of units designed to represent a particular kind
of information (e.g., verbal vs. visual) or to accomplish a par-
ticular information processing goal (e.g., input vs. output).
For example, Farah and McClelland’s (1991) model of
semantic memory impairment has three modules correspond-
ing to verbal inputs, to visual inputs, and to semantic repre-
sentations (which are further subdivided into visual units and
functional units). Units within a module are richly intercon-
nected with each other, and units in different modules may or
may not be connected depending on the architecture of
the model. For example, in Farah and McClelland’s model,
visual input units and verbal input units are connected to se-
mantic representation units but not to each other.
Presenting a stimulus to the network causes an initial pat-
tern of activation across the units, with some units more ac-
tive than others. This pattern changes as each unit receives
activation from the other units to which it is connected. A sta-
ble pattern of activation eventually appears across the units.
The particular pattern instantiated across a set of units in re-
sponse to an input, such as seeing an object or hearing a word,
is determined by the weights on the connections between the
units. Knowledge is therefore encoded in the weights, which
constitute the long-term memory of the network.
The feature of distributed network models that may ex-
plain more than any other their continuing influence is that
they learn. A network can be trained to produce a particular
output, such as the meaning of a word, in response to a par-
ticular input, such as the orthographic pattern of the word.
Training involves incrementally adjusting the weights be-
tween units so as to improve the ability of the network to pro-
duce the appropriate output in response to an input.
Another important characteristic of distributed network
models is that their performance can decay gracefully with
damage to the network. This characteristic is a result of hav-
ing knowledge distributed across many connection weights in
the network. For example, even with up to 40% of its visual
semantic memory units destroyed, Farah and McClelland’s
(1991) model was able to correctly associate names and pic-
tures more than 85% of the time.
Distributed network models have been applied to many
human behaviors that depend on information traditionally rep-
resented in semantic memory, including acquisition of generic
knowledge from specific experiences (e.g., McClelland &
Rumelhart, 1985), word naming and lexical decision (e.g.,
Kawamoto, Farrar, & Kello, 1994; Seidenberg & McClelland,
1989), impairments in reading and the use of meaning after
brain damage (e.g., Farah & McClelland, 1991; Hinton &
Shallice, 1991; Plaut, McClelland, Seidenberg, & Patterson,
1996), and (as discussed later) semantic priming. Although
these models have had their critics (e.g., Besner, Twilley,
McCann, & Seergobin, 1990; Fodor & Pylyshyn, 1988), their
influence on the science of memory has been, and promises to
remain, enormous.High-Dimensional Spatial ModelsThe idea that concepts can be represented as points in space,
such that the dimensions of the space correspond to important
dimensions of meaning, has a long history (e.g., Osgood,
Suci, & Tannenbaum, 1957). This idea has recently been res-
urrected in two models of the acquisition and representation
of word meaning.Hyperspace Analog to Language (HAL). HAL (e.g.,
Burgess & Lund, 2000) is a spatial model of meaning repre-
sentation in which concepts are represented as points in a
very high dimensional space. The semantic similarity be-
tween concepts is represented by the distance between corre-
sponding points in the space. As a result of the methodology
used, meanings of concepts are represented in terms of their
relations to other concepts.
The methodology involves tracking lexical co-occurrences
within a 10-word moving window that slides across a cor-
pus of text. The corpus includes approximately 300 million
words taken from Usenet newsgroups containing English