Information Technology Reference
In-Depth Information
Frequency by Regularity
are named faster than lower frequency and inconsis-
tent words. However, frequency interacts with consis-
tency, such that the frequency effect decreases with in-
creasing consistency (e.g., highly consistent words are
pronounced at pretty much the same speed regardless
of their frequency, whereas inconsistent words depend
more on their frequency). The PMSP model shows the
appropriate naming latency effects (and see that paper
for more discussion of the empirical literature).
We assessed the extent to which our model also
showed these naming latency effects by recording the
average settling time for the words in different fre-
quency and consistency groups (table 10.5). The re-
sults are shown in figure 10.16, showing the appropriate
main effects of frequency and regularity, plus the crit-
ical interaction, whereby the most consistent words do
not exhibit a frequency effect. Although the qualitative
patterns are correct in our model, the quantitative pat-
terns differ somewhat from those produced by PMSP.
For example, exception words are differentially slower
in our model relative to the PMSP.
48
Exception
Ambiguous
Regular Inconsistent
Regular Consistent
46
44
42
Low
High
Frequency
Figure 10.16: Settling time as a function of frequency and
consistency of words. Greater frequency and consistency both
result in faster settling, but there is an interaction — frequency
does not matter for consistent words.
This captures pattern
seen in human reading latencies.
was particularly true for the Glushko (1979) exception
list (for the network and for people); table 10.6 lists the
original “raw” performance and the performance where
alternate pronunciations are allowed. Also, the Mc-
Cann and Besner (1987) lists contain two words that
have a “j” in the coda, which never occurs in the train-
ingset. ThesewordswereexcludedbyPMSP,andwe
discount them here too. Nevertheless, the network did
sometimes get these words correct, though not on the
specific testing trial reported here.
Go to the PDP++Root window. To continue on to
the next simulation, close this project first by selecting
.projects/Remove/Project_0 . Or, if you wish to
stop now, quit by selecting Object/Quit .
10.4.3
Summary and Discussion
This model has expanded and elaborated the direct route
between orthography and phonology explored previ-
ously in the simpler model from section 10.3. Specifi-
cally, we have seen how the complex regularities of the
spelling to sound mapping in English can be effectively
represented using the same representational principles
as the invariant object recognition model from chap-
ter 8. Thus, the model places visual word processing
within the larger context of visual object recognition.
This model used a single pathway for processing both
regular and exception words, in direct contradiction to
the traditional dual-route theories. These theories hold
that the direct mapping between spelling and sound
should only occur for regular words via the application
of explicit rules. Nevertheless, this model exhibits sig-
nificantly poorer performance for low frequency irreg-
ulars than other word categories.
Question 10.8 Can you explain why the present model
was sometimes able to pronounce the “j” in the coda
correctly, even though none of the training words had a
“j” there? (Hint: Think about the effect of translating
words over different positions in the input.)
Naming Latencies
One final aspect of the model that bears on empirical
data is its ability to simulate naming latencies as a func-
tion of different word features. The features of interest
are word frequency and consistency (as enumerated in
table 10.5). The empirical data shows that, as one might
expect, higher frequency and more consistent words
Thus, if this model
Search WWH ::




Custom Search