Language - Computational Explorations in Cognitive Neuroscience - page 321

Phon Examples F/B U/D Rnd/Flt Sht/Lng

Front Back

123 4567

/E/ keep, heat 1 1 flat long

1
2
3

E U

/i/ hit 3 2 flat short

A

i

u

/A/ make 2 1 flat long

e Y O

o

/e/ bed, head 2 3 flat short

4
5
6

^

/@/ cat 3 5 flat short

I

@ W

/U/ boot 7 1 round long

a

/u/ put, bird 6 2 round short

/O/ hope 7 3 round long

Figure 10.3: Two-dimensional organization of the vowels
according to the position of the tongue: front or back and up
or down.

/o/ dog, call 7 4 round short

/a/ pot, father 5 6 flat short

/ ^ / cup, ton, the 5 4 flat short

/I/ bike 4 5 flat short

/W/ now, shout 6 5 flat short

is a distributed representation that roughly captures the
similarity structure among different phonemes.
The two general categories of phonemes, vowels and
consonants, have their own sets of phonological char-
acteristics that we will discuss in turn. Although there
is an international standard labeling system for the dif-
ferent phonemes (a phonetic alphabet), we will not
use these labels here because they require nonalpha-
betic symbols, and sometimes take more than one char-
acter to represent. Things are much simpler for the
models and programs that operate on them if we can
use standard English letters to represent the phonemes.
We adopt such a phonemic alphabet from Plaut, Mc-
Clelland, Seidenberg, and Patterson (1996) (hereafter
known as PMSP).

/Y/ boy, toil 5 3 round short

Table 10.1: Representations for vowels, using the PMSP
phoneme labels, with features based on location of tongue
(front/back and up/down) and the lip position (round or flat)
and short versus long. Note that the long vowels are repre-
sented by capital phoneme letters, and capitals are also used
to represent the diphthongs (/I/, /W/, /Y/).

The activity pattern corresponding to a vowel in our
network contains 1 unit active out of a group of 7 rep-
resenting front/back, 1 out of 6 for up/down, and then
1 out of 4 representing one of the four possible combi-
nations of round/flat and short/long. This coding was
designed to avoid features that are active 50 percent of
the time, which would be inconsistent with the sparse
representations the network itself develops.

Vo w e l s

Vowels form the central “core” of a word or syllable,
are always voiced, and provide a means for letting air
escape as a word is said. You can sing a vowel sound
for a while, but the same is not usually true of con-
sonants. Vowels vary along four different dimensions
as captured in our representations. Figure 10.3 shows
two of these dimensions, based on the position of the
tongue along both the front-back and up-down dimen-
sions within the mouth. The other two dimensions are
the positions of the lips (either rounded or flat) and the
“length” of the vowel sound (either short or long). A
long vowel is typically made with the vocal apparatus
more “tensed” than a similar short vowel. Table 10.1
shows each vowel with an example for you to try out,
and its values along each of these dimensions.

Consonants

Unlike the vowels, the consonants are typically pro-
duced by restricting airflow, which causes a distinc-
tive sound depending on where the restriction occurs,
the way in which it is done (called the manner ), and
whether it is voiced. For example, the /s/ sound is made
by pushing unvoiced air through a small opening pro-
duced by pushing the tongue up against the gums (also
known as the alveolar ridge ). The /z/ sound is just
the same, except that it is voiced (the vocal cords are
closed). Thus, the three critical features in our conso-
nant representations are the location at which the air-
flow is restricted, the manner, and voicing.