Language - Computational Explorations in Cognitive Neuroscience - page 341

Control Nonwords

Nonword Set Model PMSP People

Word Phon Repi Output Comment

Glushko regulars 95.3 97.7 93.8

gurst gurst gggurst gggurtt

Glushko exceptions raw 79.0 72.1 78.3

dawp dop dddoppp ddd@ppp ap ! /@p/ (lap, etc.)

phoyce fYs fffYsss fffOsss final e = long vowel

Glushko exceptions alt OK 97.6 100.0 95.9

shret Sret Srrettt Trrettt also output sret

McCann & Besner ctrls 85.9 85.0 88.6

tolph tolf tttolff tttOlTT coda ph low freq, ! th

zupe zUp zzzUppp yyyUppp

McCann & Besner homoph 92.3 n/a 94.3

snocks snaks snnakss snnakkk inflected form?

Taraban & McClelland 97.9 n/a 100.0 1

goph gaf gggafff gggappp ph low freq in coda

lokes lOks lllOkss lllekss vowel migration

broe brO brrO— brrU—

Table 10.6: Summary of nonword reading performance. The
raw values for the Glushko are for only the single provided
output, whereas alt OK shows the results when alternative out-
puts that are consistent with the training corpus are allowed.
1. Human accuracy for the Taraban & McClelland set was not
reported but was presumably near 100 percent (the focus of
the study was on reaction times).

faije fAj fffAjjj fffAzzz no err: no j in coda

zute zUt zzzUttt yyyUttt cute ! kyUt

yome yOm yyyOmmm yyy ^ mmm some ! s ^ m

zope zOp zzzOppp nnnOppp z low freq, nope

jinje jinj jjjinjj jjjInzz no err: no j in coda

Homophone Nonwords

Word Phon Repi Output Comment

stawp stop sttoppp stt@ppp

shooze SUz SSSUzzz sssOzzz e = long vowel

golph golf gggolff gggOlTT coda ph low freq, ! th

Regular Nonwords

phocks faks fffakss fffekss

Word Phon Repi Output Comment

coph kaf kkkafff kkkaSSS coda ph low freq, ! sh

mune myUn myyUnnn mmmUnnn fownd fWnd fffWndd fffXndd X = phoneme error

wosh waS wwwaSSS wwwoSSS alt: os ! /o/

bawx boks bbbokss bbbosss also output bok

Exception Nonwords waije wAj wwwAjjj wwwAzzz no err: no j in coda

muel myUl myyUlll yyyUlll also output mUl

Word Phon Repi Output Comment

binje binj bbbinjj bbbinss no err: no j in coda

blead blEd bllEddd blleddd alt: bread ! /bred/

bood bUd bbbUddd bbbuddd alt: good ! /gud/

bost bost bbbostt bbbOstt alt: host ! /hOst/

Table 10.8: Errors on the McCann & Besner (1987) control
and homophone nonwords. Columns are as in table 10.7.

cose kOz kkkOzzz kkkOsss alt: dose ! /dOs/

domb dam dddammm ddd ^ mmm alt: numb ! /d ^ m/,

some ! /s ^ m/

doot dUt dddUttt dddXttt X = phoneme error

Word Phon Repi Output Comment

grook grUk grrUkkk grrukkk alt: book ! /buk/

bood bUd bbbUddd bbbuddd alt: good ! gud wone wOn wwwOnnn www ^ nnn alt: done ! /d ^ n/

wull w ^ l www ^ lll wwwulll alt: full ! /ful/

Table 10.9: Errors on the Taraban & McClelland (1987) non-
words. Columns are as in table 10.7.

Table 10.7: Errors on the Glushko (1979) regular and ex-
ception nonwords. Phon is the phonological representation,
Repi is the repeating consonant phonology, Output is the ac-
tual output produced by the network (an X indicates uninter-
pretable phoneme output), and Comment provides explanation
of network output in terms of the training corpus, alt = valid
alternative pronunciation produced by network.

monosyllables, and applying these to nonwords.
For your convenience, the errors for all the nonword
lists are summarized in tables 10.7 through 10.9. The
full log of network responses is available in ss.resp
in the directory where the project is. To pick one re-
sponse across all the different locations of a given word,
we just picked a correct response if there was one, and
if not, we picked an informative error if there was any
variation.
As indicated in the Comment column in tables 10.7
through 10.9, we tried to determine for each error why
the network might have produced the output it did. In
many cases, this output reflected a valid pronunciation
present in the training set, but it just didn't happen to
be the pronunciation that the list-makers chose.

Continue to StepTest through some more items on
this and the other two testing lists (using PickTest to
switch to a different list).
The total percentages for both our model, PMSP
(where reported) and the comparable human data are
shown in table 10.6. Clearly, the present model is
performing at roughly the same level as both humans
and the PMSP model. Thus, we can conclude that
the network is capable of extracting the often complex
and subtle underlying regularities and subregularities
present in the mapping of spelling to sound in English

This