140
G. Sartori et al. /Neuroscience Letters 390 (2005) 139-144
of category-specificity in semantic memory. This proposal
has been enormously influential, spanning an entire area of
empirical enquiry [1,2,6,7,14,16,18,22].
At neural level, it is believed that sensory experienced
knowledge is stored in circumscribed brain regions, in a
feature-based format, which is related to the encoding sensory
channels. Functional imaging data consistent with this claim
[15] have been reported and, in addition, electrophysiological
investigations have shown that the N400, a negativity induced
by semantic incongruity, is larger for Sensory features as
compared to Non-sensory features [3]. This latter difference
has been interpreted as a neurophysiological evidence of sep-
arate encoding of Sensory and Non-sensory semantic features
in the brain.
Here we report an ERP study, in accordance with an
opposing theory about semantic features. According to this
contrasting view, semantic features are encoded in the brain
on the basis of their contribution to the meaning of a concept.
A concept may have many semantic features, although those
really useful in distinguishing it from closely related concepts
are only a few. The information content of semantic features
may be measured by semantic relevance [19,20]. Relevance
is a measure of the contribution of semantic features to the
“core” meaning of a concept. Elite few semantic features of
high relevance are sufficient for an accurate retrieval of the
target concept. In contrast, when semantic relevance is low,
retrieval is inaccurate. Among all the semantic features of a
concept those with high relevance are also critical in distin-
guishing it from other similar concepts. The following is a
case in point: (has a trunk) is a semantic feature of very high
relevance for the concept Elephant, because most subjects use
it to define Elephant, whereas very few use the same feature
to define other concepts. Instead (Has 4 legs) is a semantic
feature with low relevance for the same concept, because few
subjects use to define Elephant but do use it to define many
other concepts. When a set of semantic features is presented,
their overall relevance results from the sum of the individual
relevance values associated with each of the semantic fea-
tures. The concept with the highest summed relevance is the
one that will be retrieved. For example, the three features
(similar to a goose), (lives in ponds) and (has a beak) have
topmost relevance for Duck, followed by Swan, and then by
Ostrich (example taken from the normative data collected by
Sartori and Lombardi [19]5). The retrieved concept, given
the three features, will be Duck, because it has the highest
relevance. Hence, overall accuracy in name retrieval is poor
when concepts have low relevance, and when they have many
other semantically related concepts with which they may be
confused. It has been shown that [20]: (i) relevance is the
best predictor of naming accuracy (at least in a “naming-
to-description” task) when contrasted to a number of other
parameters of semantic features (dominance, distinctiveness)
and of the concept (e.g. Age-of-Acquisition, familiarity and
typicality), (ii) relevance is a robust measure, not significantly
influenced by the number of concepts in the database or by
sampling errors.
Here we will report an ERP study designed to address the
issue of how semantic features are coded in the brain. In this
paper we will show that: (i) low relevance descriptions have
larger N400 with respect to high relevance descriptions; (ii)
no effects of feature type arise when relevance is matched;
(iii) no differences in N400 to differing categories of Living
and Non-living concepts can be detected when relevance is
matched.
Twenty-four Italian undergraduate students (age range
19-29 years; mean = 22.6, S.D. = 2.55) participated in the
experiment. Nine were male and 15 female. Average edu-
cation was 16.7 years. All the subjects were healthy and had
normal or corrected-to-normal vision.
Every trial consisted in the sequential presentation of a
verbal description of three semantic features on a computer
screen (e.g. (has a carriage), (found in the airport) and (found
in the sky)) followed by the presentation of a target word
(e.g. Airplane) after which a Yes/No response was required.
The task was to indicate whether the three features correctly
indexed the concept or not. Half subjects responded with their
right hand using the index finger for Yes responses and the
middle finger for No responses; the remaining half used the
fingers in opposite order.
In regard to the experimental stimuli, they varied accord-
ing to the following dimensions: (i) Category (Living versus
Non-living); (ii) Relevance (High versus Low); (iii) Feature
type (Sensory versus Non-sensory); (iv) Congruency (Yes
versus No). A total of80 concepts were used. For each con-
cept four descriptions were presented (two of high relevance,
one Sensory and one Non-sensory and two low relevance,
again one Sensory and one Non-sensory). These 320 stim-
uli were followed by the target concept and required a Yes
response. Target words were matched across categories (Liv-
ing n=40 and Non-living n= 40) for Age of Acquisition
(p = 0.58), Typicality (p = 0.90) and Familiarity (p = 0.60)
(norms collected by Dell’Aqua et al. [5]). Average seman-
tic relevance for Living (2.73) did not differ from that of
Non-living (2.83) (p = 0.51). Average semantic relevance for
Sensory features (2.80) did not differ from that of Non-
sensory features (2.75) (p = 0.74). Relevance values of the
three semantic features presented sequentially to the sub-
jects were taken from the norms collected by Sartori and
Lombardi [19]. All the 320 stimuli requiring a No response
had the same level of dissimilarity with the correct target as
measured by standardized cosine.6 The following is a telling
example: if the correct description for the concept Peach is,
instead, followed by Violet a No response is required. The
cosine similarity of Violet with respect to Peach is 0.073
5 Relevance values are derived algorithmically from a feature-listing task
and are not based on subjective ratings. The computation is based on the
number of times people report a given feature in defining a concept [20].
6 Standardized cosine is a popular measure of similarity between vectors
of semantic features. Matching cosine similarity guarantees that the foils are
equally dissimilar to the target.