Statistical learning of conditional

orthographic correspondences
Brett Kessler
Washington University in St. Louis, MO, USA
Abstract
The English writing system deviates widely from the alphabetic ideal of uniform
one-to-one correspondence between graphemes and phonemes, but its inconsis-
Correspondence: tency is greatly reduced when conditional sound–spelling rules are applied. When
Brett Kessler,
Psychology Department,
reading or writing one part of a word, children and adults evinced knowledge of
Washington University rules sensitive to the identity of other letters or phonemes, even those appearing
in St. Louis, much later in the word. Adults also showed sensitivity to the distinction between
Campus Box 1125, the basic and Romance subsystems of English (Albrow’s Systems 1 and 2).
One Brookings Drive, Children as young as 6 years applied conditional rules that they were not taught,
St. Louis, MO 63130-4899,
indicating implicit statistical learning of patterns observed in text. But learning is

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

USA.
E-mail: imperfect, and even adults did not match the frequency with which the patterns are
bkessler@wustl.edu found in English words.

The English writing system is routinely criti- one-offers like ‘hiccough’, where <gh> = /p/, and
cized for falling far short of the alphabetic ideal ‘of’, where <f> = /v/.1 It also provides a straight-
(Caravolas, 2004) of one-to-one correspondence forward explanation of how children can learn to
between phonemes and graphemes. Almost all its read and write in classrooms whose curriculum
phonemes have multiple spellings—sound-to-letter, is based on whole-word learning with no expla-
or spelling inconsistency—and almost all its let- nation of how sound–spelling correspondences
ters have multiple pronunciations—letter-to-sound, work. But there is also much evidence that
or reading inconsistency. A study by Hanna et al. whole-word memorization cannot be the whole
(1966) showed that only 73% of all phonemes picture. All other factors being equal, people are
would be spelt correctly if the writer picked the most slower and more error-prone at processing words
common spelling for each phoneme: on average, with unusual sound–spelling correspondences
a person who spells by the alphabetic principle would (Balota et al., 2004), which should not be a fac-
make a mistake on every fourth grapheme. tor if they memorized all words as arbitrary letter
But arguably the speller’s loss is the psycholin- sequences. Even more convincing is the fact that
guist’s gain. The fact that most mature spellers do people can come up with plausible spellings when
much better than 73% raises questions as to how asked to write non-words, and other people can
people read and write words, if applying one-to-one read such spellings. These feats would hardly be
correspondences is not enough. Some theories that possible if the only route to reading and writing
have been adduced are: were memorizing words as unanalysable wholes.
• Learning large-unit sound–spelling correspon-
• Whole-word memorization: people memorize dences: it has often been noted that the rimes
entire words as arbitrary sequences of letters. (the vowel plus the following, coda, consonants)
Such a theory accounts for the fact that people of riming words are often spelt alike. There may
learn all sorts of inconsistent spellings, including be many ways to spell /aɪ/ in English, but the

Writing Systems Research, Vol. 1, No. 1, 2009. © The Author 2009. Published by Oxford University Press. 19
For Permissions, please email: journals.permissions@oxfordjournals.org
doi:10.1093/wsr/wsp004
B. Kessler

spelling of the rime /aɪld/ is consistently <ild>, chess: a set of linguistically encoded instructions
at least in simple monomorphemic words like that one consciously learns from a teacher and can
‘wild’ and ‘child’. The idea that people might easily formulate and pass on to one’s student. Even
learn the spelling of whole rimes as units is spelling researchers may have a hard time explaining
attractive because there is much evidence that that /k/ is spelt <ck> at the end of word-final stressed
preliterate children are good at breaking syllables syllables when the rime of the uninflected lexeme
down into onsets (the part before the vowel) and would otherwise contain fewer than three letters
rimes—while breaking rimes down into vowels (e.g. ‘sick’ versus ‘silk’, as discussed below), and
and codas is often very challenging for them most people who follow the rule have never been
(Goswami, 1993; Treiman and Kessler, 1995). taught it overtly (Hayes et al., 2006). An alternative
Much work in literacy research has proceeded hypothesis is that complex sound–spelling corre-
from the point of view that onsets and rimes spondences are learnt the same way we learn many
are the effective units of reading and writing other patterns in life: by observing and internalizing
(Stanback, 1992; Ziegler et al., 1997), which the relative frequency with which objects and events
is not to claim that these researchers subscribe occur and co-occur. Perhaps through repeated expo-
to the theory that literacy is nothing more that sure to text, people gradually pick up both uncon-
memorizing the spellings of whole onsets and ditional and conditional correspondences, without
rimes. One problem with such a theory is that necessarily formulating any conscious accounts of

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

many rimes have multiple spellings (‘bed’, ‘said’, them. This type of learning can be called statisti-
‘head’). Furthermore, Aronoff and Koch (1996) cal learning, in the sense that it is based on implicit
reported that there are only 12 rime spellings in numeric analysis: orthographic patterns are essen-
English that are worth memorizing as wholes. tially observations (or computations) that particular
• Learning conditional sound–spelling correspon- sound–spelling correspondences are more frequent
dences: although the alphabetic ideal is typically than others in a particular context. It is important
described as having only unconditional rules to keep in mind that most vocabulary about mental
like ‘the letter <n> spells the phoneme /n/’, processes—words such as ‘rules’, ‘observation’,
full stop, researchers such as Carney (1994), ‘computation’—carries an implication of conscious,
Cummings (1988), and Venezky (1970, 1999) explicit thought, but in the context of this article I do
have shown that many correspondences have not intend any such implication.
fewer exceptions when stated as rules that specify With this proviso in mind, I address in this article
conditions—rules like ‘<n> spells /ŋ/ before <k>’. such questions as whether people notice and use
Conditional spelling rules can, in principle, refer spelling patterns they have never been taught, and if
not just to phonetic environments but also other so, how. Are people deterministic computers whose
word properties such as morpheme class, word knowledge of letter frequencies perfectly reflects
length, or etymological stratum: ‘<ph> spells /f/ their experience? At what age do people begin to
in words of Greek extraction’. Indeed, even one- build up this knowledge? Do they use it for both
off irregularities can be formulated as conditional reading and for writing? And what types of patterns
rules: ‘<gh> spells /p/ in the word “hiccough” ’. can people pick up on?
Thus, conditional rules of orthography are very
powerful but also potentially complex.
1 Vocabulary Statistics
In this article, I report some recent psycholin-
guistic work in which my colleagues and I have A first step in exploring this statistical learning
explored whether people learn and use conditional hypothesis was to see what sort of usable spelling
sound–spelling rules, and if so, how. I do not mean patterns could be found in English words. Kessler
that these are necessarily rules in the everyday sense and Treiman (2001) were primarily interested in
of the term, as when one talks about the rules of learning how the predictability of sound–spelling

20 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

correspondences in one part of the syllable is Table 1 Sound–spelling consistencies of the parts of
improved when rules are sensitive to the contents monosyllabic words
of other parts of the syllable. Because the syllabifi-
Contexta Onset Vowel Coda
cation of intervocalic consonants is a moot issue in
English, we found it expedient to restrict this first Reading b

study to monosyllabic words. We gathered 3,117 Unconditional 0.976 0.717 0.982

words, basing our selection primarily on familiar- Given onset – 0.807 0.992
Given vowel 0.993∗ – 0.992
ity ratings obtained from American college stu- Given coda 0.988 0.920∗ –
dents (Nusbaum et al., 1984). Pronunciations were Writingc
taken from the Random House Dictionary (Flexner, Unconditional 0.910 0.529 0.821
1987). When multiple pronunciations were listed, Given onset – 0.649∗ 0.882
we selected the first general American pronuncia- Given vowel 0.937∗ – 0.925∗
tion, but homographic heterophones such as ‘read’ Given coda 0.942 0.737∗ –
(/ɹid/, /ɹɛd/) were treated as separate words. After a
‘Unconditional’ is unconditional consistency; the others are
splitting the pronunciation into onset, vowel, conditional, with the context noted after the word ‘Given’.
b
and coda, we aligned those three parts with their Letter-to-sound consistency.
spelling. For example, the word ‘crane’ aligned like c
Sound-to-letter consistency.
this:<cr> = /kɹ/, <a_e> = /e/, <n> = /n/. ‘Sign’ aligned ∗
Consistency gain over unconditional value is significantly greater

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

<s> = /s/, <ig> = /aɪ/, <n> = /n/, because we adopted than when the values in the context are randomized, P < 0.001.
Albrow’s (1972) criterion that all letters must align
with some phoneme.
With all these alignments in place, we were in the cell labelled ‘Onset Reading Given vowel’
able to compute consistency measures for the vari- that the letter-to-sound consistency of the onset rises
ous orthographic units in the three parts of the syl- to an almost perfect 99.3% when the vowel is taken
lable. For example, in the onset of the words, the into account. Similarly, all the other conditional
letter <c> has a consistency of 88.4%, because consistencies are higher than the unconditional con-
of the ninety-seven words that begin with <c>, sistencies. It is particularly noteworthy how much
ninety-one (93.8%) are pronounced with /k/ and more reliably one can read vowels when the coda is
the remaining six (6.2%) are pronounced with /s/; taken into account, and how much more reliably one
the weighted average of those two proportions is can spell all parts of a syllable when the identity of
88.4% = (91 × 93.8% + 6 × 6.2%) ÷ 97. (All counts are the adjacent syllable position is taken into account.
by word types, not weighted by their frequency in The mere fact that there are increases is not at all
running text.) Further, by averaging such values for surprising: extra information can never hurt, even
all letters in syllable onsets, we concluded that the when the associated values are purely coincidental.
average consistency of letter-to-sound mappings in In fact, in many cases it turned out that the condi-
onsets was 97.6%. The rows labelled ‘Unconditional’ tional consistencies were comparably high even if
in Table 1 report these unconditional consistency we scrambled the contexts, arbitrarily swapping,
measures, as proportions ranging from 0 to 1, for for example, the coda of each word with the coda
all the three parts of the syllables. They quantify the of another. But for some contexts, the ones denoted
impression that vowels are more difficult than con- by asterisk in Table 1, the attested conditional con-
sonants, and that writing is harder than reading. sistencies were significantly higher than when the
Measurements of conditional correspondences contexts were randomized. These contexts tend to
were more directly relevant for our purposes. For be more systematic, typically raising the consisten-
example, we computed the consistency of initial cies because of patterns that are due to conditional
<c> when the vowel is <a>, its consistency when the sound changes in the history of the language. For
vowel is <i>, and so forth for all onsets and all vow- example, <i> is particularly likely to be pronounced
els, then took the weighted average. Table 1 shows /aɪ/ when the coda is <ld> because of a lengthening

Writing Systems Research, Vol. 1, No. 1, 2009 21

B. Kessler

that occurred in Old English, and English has had 2 Spelling Vowels in
many other vowel changes that were conditioned by
specific codas. In contrast, very few sound changes
Non-words Using Conditional
in the onset were conditioned by the coda, or vice Correspondences
versa, which accounts for the non-significant con-
nections between those syllable parts. The next question is whether humans are sensitive
It bears remembering that these conditional to, learn, and use conditional patterns. Treiman
consistencies were all derived by mathematically et al. (2002) explored how people spell vowels,
analysing the vocabulary itself: the data do not say which, as Table 1 shows, have the most to gain
that people take advantage of these patterns, but from conditional reference to context. Do adults
they do prove that the patterns are there (a full list take adjacent consonants into account when they
of the individual patterns can be found at http://spell spell vowels? We selected nine vowels for which
.psychology.wustl.edu/RelSoundLetMono). The data our vocabulary analysis had revealed that a given
also give us some basis for addressing the large-unit spelling was much more likely in one context than
hypothesis raised earlier: would it make sense to another. These are listed in Table 2. For instance,
memorize the spellings of entire onsets and rimes? it turns out that stressed /i/ is spelt <ee> in 64% of
The fact that the consistencies of vowels are so much all words where it appears before word-final /d/ or
higher when the coda is taken into account, and vice /p/, but that <ee> is used only 15% of the time, on

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

versa, means that such a strategy would be much average, before several other consonants, such as
better than only learning unconditional phoneme- /ð/ and /m/. The difference between those two num-
level correspondences. On the other hand, the data bers is presented as a proportion (0.49) in the table.
also show that taking the vowel into account can There are more words like ‘bleed’ than ‘mead’ and
help one to read and spell the onset, and taking the more words like ‘dream’ than like ‘seem’, but even
onset into account can help one to spell the vowel. the minority spellings are not particularly unusual.
Just memorizing large units like onsets and rimes Other patterns that were tested were less violable.
in an unconditional manner would ignore much For instance, /aʊ/ is quite reliably spelt as <ow>
information that could improve reading and spelling before coda /l/ and /n/ (‘clown’), although even in
accuracy. those contexts there are exceptions, such as ‘noun’.

Table 2 Increase in proportion of vowels that adults spelt with the conditioned spelling when a conditioning consonant
is present

Vowel Conditioned spelling Context Example Vocabularya Responsesb

Before coda
/ε/ <ea> /d/ head 0.40 0.06
/i/ <ee> /d/, /p/ feed 0.49 0.27
/o/ <o> without <e> /l/ roll 0.33 0.11
/ʊ/ <u> /l/, /s/, /ʃ/, /tʃ/ push 0.73 0.11
/aʊ/ <ow> /l/, /n/ clown 0.89 0.43
/aɪ/ <igh> /t/ night 0.39 0.27
After onset
/ɑ/ <a> /w/ wasp 0.75 0.72
/ɚ/ <or> /w/ word 0.52 0.28
/u/ <oo> Non-coronals food 0.57 0.58
a
Based on standard spellings at the end or beginning of English words. Proportion of words with indicated vowel and consonant that
have the indicated spelling, minus proportion of words with the vowel and another consonant that have the indicated spelling.
b
Results from experimental trials, measured the same way as the vocabulary statistics

22 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

Three of the patterns involved the influence of the /ɛ/ = <ea> before /d/ could be due to a different
onset. For example, we set up stimuli that tested mechanism whereby spellers memorize spellings for
whether people were sensitive to the pattern by which whole rimes (/ɛd/ = <ead>) and onsets. In this experi-
General American /ɑ/ (this would be /ɒ/ in some ment, that possibility is contradicted by the fact that
other accents) is spelt as <a> 86% of the time it the same spellers who drew on the identity of the
appears after /w/ (‘wand’), but only 11% of the coda to spell vowels (intra-rime effects) also took
time after other consonants (not ‘∗fand’ but ‘fond’; into account the identity of the onset. For example,
the difference 0.75 appears in the table). Thus, we they spelt /ɑ/ as <a> the great majority of the time
investigated a wide variety of patterns. after /w/, but comparatively rarely after other conso-
To test whether spellers were sensitive to a par- nants. Thus, a whole-rime spelling strategy appears
ticular pattern, we designed ten pairs of non-words to be not only an insufficient explanation, but also an
like /θεd/ and /θεk/ and asked adults to spell them unnecessary one: the same type of (implicit) aware-
as if they were English. The pairs were split up and ness of conditional patterns that is needed to explain
interleaved with similar questions for the other eight the influence of the onset can be used to explain the
patterns, so that the participants would not deduce influence of the coda.
any patterns from the stimuli in the experiment itself.
In this particular case, we tallied whether they used
<ea> significantly more often in non-words like /θεd/ 3 Conditional Spelling of

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

than in those like /θεk/. And that proved to be true. In Vowels in Real Words
this case, the percentage of non-words like /θεd/ that
were spelt with <ea> was 11%, whereas only 5% of Non-words can be ideal stimuli for orthographic
the stimuli like /θεk/ had <ea>; the difference of 6% is experiments, because they do not carry the same sort
what is reported in the ‘Responses’ column in Table 2. of baggage that words do. Factors such as familiar-
This is a fairly small effect, but it was statistically sig- ity, frequency, age of learning, and semantic factors
nificant (the cutoff P < 0.05 will be used throughout such as imageability can differentially affect how
this article) and, as the table shows, the other cases we people read and spell words (Balota et al., 2004).
tested had effects of larger magnitudes. There is always the concern that when we think
Thus, adult spellers have learnt conditional pat- people are reacting distinctively to a group of words
terns, and they do make active use of them in their because of one property, such as high inconsistency,
own productions. This knowledge is statistical in they are really reacting to some other property, such
several senses of the word. The patterns are rarely as lower familiarity, which we were unable to con-
if ever taught explicitly or recalled consciously, but trol for or might not even have known about. In the
appear to be the product of long experience seeing previous experiment, if people were asked to spell
different vowel spellings being used in different con- real words, such as ‘dead’ versus ‘deck’, the finding
texts. They are not exceptionless rules but statements that people used more <ea> before /d/ would be com-
of frequency distributions. Indeed, a few of the rules promised by the obvious objection that the results
are, at least by themselves, rather useless in spelling: could be explained by whole-word memorization.
it is never advisable to guess that the spelling of /ɛ/ Because we used non-words such as /θεd/ versus
is <ea>, because even before /d/ that spelling is less /θεk/, which people had never seen spelt before,
frequent than <e> (‘bed’). Nevertheless, <ea> is in such an objection is untenable, and the implication
the speller’s repertoire, and it has a higher probabil- that the participants used conditional patterns is
ity of being selected before some codas than before much more believable.
others. The experiment demonstrated that statisti- Nevertheless, there is a sense in which studying
cal patterns of conditional probabilities do play an the spelling of words via non-words is somewhat
important role in phonological spelling. artificial. One might reasonably object that non-
The experiment also answered the question as word spelling is an imperfect window on natural
to whether apparently conditional spellings like spelling, because people might write them using

Writing Systems Research, Vol. 1, No. 1, 2009 23

B. Kessler

different strategies than when writing familiar mind is that we intentionally selected the critical
words. Consequently, we configured another ver- stimuli from the minority of words that go against the
sion of the experiment to use actual words (Treiman general pattern. Even if sensitivity to context leads
et al., 2002). We asked the participants to spell pairs some spellers astray on a few words like ‘wombat’,
like ‘shred’ and ‘fleck’. We then performed error that same sensitivity should help them out on the
analyses to see how often they misspelt the vowel much larger number of words in which /ɑ/ is spelt
as <ea>, testing the hypothesis that such an error <a> after <w>, such as ‘waffle’ and ‘wallow’.
would be more likely in the context (coda /d/) that
is more strongly associated with the /ε/ = <ea> corre-
spondence in English orthography. Other examples 4 Reading Vowels Using
included looking for <ee> in ‘reap’ versus ‘ream’; Conditional Rules
<o> without silent <e> in ‘shoal’ versus ‘croak’; <u>
in ‘swoosh’ versus ‘nook’; <ow> in ‘pronoun’ versus Much more research has studied how people read
‘slouch’; <igh> in ‘contrite’ versus ‘confide’; <a> than how they spell. It has often been reported that
in ‘wombat’ versus ‘possum’; <o> in ‘whir’ versus readers’ pronunciation of vowels varies depend-
‘blurt’; and <oo> in ‘scuba’ versus ‘frugal’. Because ing on the coda (e.g. Johnson and Venezky, 1976;
errors of any type would be more numerous on lon- Glushko, 1979; Ryder and Pearson, 1980; Andrews
ger or less familiar words, the stimuli were matched and Scarratt, 1998). However, the prior literature left

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

for length and familiarity. An attempt was made to several issues open. As discussed above in the case
select words that college students would know, but of spelling, one interpretation of a coda effect is that
not so well that they would spell them perfectly. people memorize letter-to-sound correspondences
The errors went in the expected direction: 27% at the level of whole onsets and rimes: if a person
of the errors in the conditioning context constituted reads <nind> as /naɪnd/, perhaps they have broken
use of the conditioned vowel, but only 8% of the the word down into <n> and <ind>, and applied the
errors in the other contexts were of that nature. That correspondences <n> = /n/ and <ind> = /aɪnd/, treat-
is, we often saw errors like <reep> instead of <reap>, ing the whole rime as a unit. A rime-unit strategy
but much less often did we get errors like <reem> could be distinguished from a conditional strategy
for <ream>. The only tested pattern that was not (<i> = /aɪ/ before <nd>) if vowel pronunciations were
reflected in the participants’ misspellings was the use found to be influenced by onset consonants as well
of <u> before /ʃ/, as in ‘push’, and that may be due as by coda consonants. Several studies (e.g. Treiman
to the fact that the only item tested, ‘swoosh’, would and Zukowski, 1988; Treiman et al., 1995; Andrews
have unusual graphotactics, the sequence <wu>, if and Scarratt, 1998) reported that the onset has negli-
the contextual rule were applied to it. gible influence on vowel reading, thus leaving open
Thus, people do use statistical conditional rules the possibility that readers do follow a rime-unit
in a naturalistic spelling task, when trying to recall strategy.
the spelling of moderately familiar words that they Treiman et al. (2003) addressed this issue with an
do not encounter every day. Granted, the way the experiment structurally similar to that of the spell-
experiment was framed, it sounds like context sen- ing experiment reported earlier, but geared towards
sitivity is a bad thing, because we investigated how conditional patterns that are effective in the reading
contexts like /p/ after /i/ could mislead spellers into direction. Adults were asked to read, interleaved
making errors. One must keep in mind, however, among other trials, pairs of non-words such as
that more errors on words like ‘reap’ mean fewer <spange> and <spance>. If they pronounced words
errors on words like ‘ream’. It is easy to discount of the former type with /e/ more than they did the
/m/ after /i/ as being a conditioning context, because latter, that would show sensitivity to the fact that the
the spelling it conditions, <ea>, is the more common coda spelling <nge> conditions the /e/ pronunciation
one, but from a statistical perspective /m/ is just as (words like ‘range’) whereas <nce> does not (words
much a context as /p/ is. Another thing to keep in like ‘prance’). Two of the patterns tested included

24 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

onset conditioning. Table 3 shows the patterns that and because the great majority of contextual effects
were tested, along with their reliability in text, on vowel pronunciation come from within the rime
measured as in Table 2. Most of the patterns were (i.e. the coda), one might expect that people would
quite strong, but not completely exceptionless (e.g. be locked in to look for patterns within rimes, or
‘flange’ is pronounced with /æ/). at least learn and apply them much more robustly.
The last column in Table 3 shows that the readers Instead, it appears that people learn statistical pat-
were sensitive to the patterns. The numbers are the terns wherever they exist.
difference between the proportion of pronunciations
in the conditioning context (e.g. <nge>) that have the
conditioned vowel (/e/) and the proportion of those in 5 Magnitudes and Computer
the non-conditioning context (<nce>) that have that Modelling
vowel. Thus, any positive proportion in this column
shows a trend towards sensitivity to the context; its A curious fact about the experiments with non-word
maximum is 1.0. In all cases, the differences between stimuli is that the participants’ rate of use of statisti-
the two contexts were statistically significant. cal patterns was lower than that found in the gen-
It is important to note that the two sets of items eral vocabulary. For example, while 40% of English
that involved an onset context had just as strong words ending in stressed /ɛd/ are spelt with <ea>, the
a response as the items that had only a coda con- participants in our experiment used <ea> for only

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

text. This similarity is expected under the theory 11% of the non-words that ended in /ɛd/ (Table 2).
that people learn conditional patterns, but it would Across both experiments, participants approached
not be expected if people rely on rime-unit reading. the frequency attested in the lexicon in only three
In fairness to previous researchers who found no of the seventeen patterns we examined. If readers
reliable effect of onset, there are few cases in English and writers were perfect learners of the statistics of
where the pronunciation of a vowel letter is strongly the writing system, and if the process of spelling
sensitive to the identity of the onset; indeed, the only consisted entirely of retrieving the most probable
truly important cases may be those involving /w/. relevant patterns, we would expect the attested fre-
Thus, it is true that the overall impact of the onset is quencies to be closer to those that people observe in
small. But from a psychological point of view, that text (Brown, 1998). Although we have successfully
makes it all the more striking that people pick up on predicted that people will learn and apply certain pat-
the few onset–vowel patterns that do exist. Because terns, we have no good model for predicting the mag-
rimes are particularly salient for English speakers, nitude with which they will apply those patterns.

Table 3 Increase in proportion of vowels read with conditioned pronunciation when a conditioning consonant is present

Conditioned
Vowel Context Example Vocabulary Responses
pronunciation

Before coda
<a> /e/ <nge> range 0.95 0.54
<a> /ɔ/ <ld>, <lt> bald 0.79 0.86
<ea> /ɛ/ <d> dread 0.78 0.12
<i> /aɪ/ <nd>, <ld> find 0.76 0.33
<o> /o/ <ld>, <lt> bold 0.96 0.83
<oo> /ʊ/ <k> look 0.96 0.70
After onset
<a> /ɑ/ <u>, <w> wand 0.93 0.58
<a> /ɔ/ <u>, <w>, before <r> warm 0.96 0.16
Numbers are differences between proportions, analogous to those in Table 2.

Writing Systems Research, Vol. 1, No. 1, 2009 25

B. Kessler

It is always a challenge to predict how strongly do a good job simulating human reading of vowels
people will react to environmental stimuli, and I will whose pronunciation is conditional, and this is partly
not conclude this section with a bold new model true. Like our research participants, the connection-
that accounts for why the vocabulary numbers are ist models did learn the conditional patterns, and
so much greater than the participants’. However, they did tend to use the conditioned pronunciations
it was instructive to look at computer models of more in the strongly conditioning environments than
reading to see how their numbers compared with in other environments—e.g. they pronounced <a> as
those of the vocabulary and of our participants. We /e/ more before <nge> than before <nce>. They also
examined the output of twelve different simula- matched human performance quite closely on non-
tors and looked for implementation characteristics words that have environments that only weakly con-
that provided the closest match with the human dition pronunciations, e.g. stimuli that end in <nce>,
behaviour (Treiman et al., 2003). with agreement levels as high as 94–96% for some
The most prominent of the current spelling models (those of Plaut et al., 1996; Zorzi et al., 1998;
models is DRC (the dual-route, cascaded, model of Powell et al., 2001; Harm and Seidenberg, 2004).
Coltheart et al., 2001). This model pronounces non- These are all notable achievements. However, the
words by applying a series of letter-to-sound rules connectionist models did not match the magnitudes
that are almost all unconditional. Not surprisingly, of human performance very closely on the strongly
DRC scored 0.00 on all patterns: that is, it was no conditioning environments like <a> before <nge>.

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

more likely to use a conditioned vowel pronuncia- The best performance on such non-words was by
tion in a conditioning consonant context than when the model of Norris (1994), which matched human
the vowel letter appeared in another context. Clearly, performance only 68% of the time; some other
unconditional rules are not a good model of human models performed as badly as 43% on such stimuli
behaviour for these stimuli. (Zorzi et al., 1998).
Most other reading models pronounce non- A closer error analysis reveals a few reasons
words by applying connectionist principles. Broadly for the models’ performance. The model of Zorzi
speaking, a connectionist model simulates a neu- et al. (1998) was intentionally designed to severely
ral network: it has units that represent neurons and restrict the complexity of patterns that the system
connections between units that represent synapses. could learn, in that it lacked hidden units and its
Processing, such as reading, is a matter of setting the connections were set up so that patterns could only
activation levels of a certain set of units to represent be learnt between vowels and codas, not between
the input (letters); spreading the activation through vowels and onsets. At the other extreme, the model
the net along the connections between the units; then of Norris (1994) specialized in contextual effects:
reading off the activation levels at the set of units vowels were set up to be read in a way that explic-
that represent the output (phonemes). Learning is itly took into account how they were pronounced
a matter of adjusting the strengths of connections in words that shared the same onset or coda. This
until the processing of a training set—in our case, strategy worked moderately though not spectacularly
a word list containing spellings and pronunciations— well for non-words that have a strongly conditioning
is optimized (Rumelhart and McClelland, 1986). onset or coda, but the strategy apparently only got
At no point do classical connectionist models for- in the way of reading non-words whose vowel has
mulate explicit rules or even store specific frequency a typical unconditional pronunciation. Most of the
information at a particular locus: information is dis- remaining models applied the conditioned pronunci-
tributed throughout the network of units. ations in conditioning environments (e.g. using /e/ in
In many respects, the design of connection- <a> + <nge> words) more than the humans did, often
ist models is redolent of the way people gradually at rates approaching or even exceeding the patterns’
and implicitly learn spelling patterns through long frequency in actual vocabulary.
exposure to and use of the spelling system. We A couple of models (Plaut and McClelland, 1993;
might, therefore, expect connectionist models to Powell et al., 2001) were trained not only on English

26 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

vocabulary, but also on individual letter-to-sound But literacy researchers pay at least as much atten-
rules like <a> = /æ/, because many children are heav- tion to how children read and write words. Children
ily exposed to these unconditional rules in school, who have not yet mastered a writing system are the
and so may tend to over-apply them, therefore best sources of information on how literacy skills are
under-applying conditional rules. Although that is acquired. There is also an important practical benefit
an ingenious idea, it did not have a noticeable effect to understanding the development of orthographic
in bringing down the programs’ rate of application knowledge, in that educators can take into account
to human levels. what sort of information children are capable of
The models also differed in many other details, dealing with at various ages, and in what order they
any of which could account for substantial differences naturally acquire different types of knowledge. See,
among their outputs. Therefore, it is perhaps best to for example, Cook (2004) for an introduction to the
think of their outputs not as rigorous tests of any par- voluminous literature on theories of literacy acqui-
ticular theory of statistical learning, but as sources sition. With respect to the question of conditional
of ideas for future research. Norris’s (1984) way of orthographic patterns, the central concerns here are
incorporating context seems promising. Instead of how old children are when they acquire sensitivity to
learning only the highly discriminating contextual context; at what rate their application of conditional
effects that we tested on—and which most connec- rules increases over the course of their life; whether
tionist models would learn most thoroughly—his they have conditional rules properly speaking, or

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

model also explicitly uses information about non- whether they use rime-unit strategies; and whether
discriminating contexts. For example, in reading some types of sensitivity to context are acquired
a non-word like <chead>, we have only been consid- earlier than others.
ering the fact that when <d> is present, the probabil- Treiman et al. (2006) replicated the above study
ity of the correspondence <ea> = /ɛ/ jumps to 78% of vowel reading, but instead of recruiting college-
(it normally spells /i/ in other contexts). We, like age adults, we studied children and young adults
most connectionist models, have been ignoring ranging in age from 6 to 17 years. Because children
the onset <ch>, because it has no effect on how vary appreciably in their rate of development, we
the following <ea> is pronounced; in ‘cheat’, for analysed the results as a function of the children’s
example, <ea> has the same pronunciation it has in grade-equivalent reading ability as determined by
the vast majority of words, /i/. However, if humans a standardized test (Wilkinson, 1993). We looked at
do learn that <ea> is always pronounced /i/ after how often they produced conditioned vowel pronun-
<ch>—a useless but true fact—and if they apply ciations when they read vowel letters in condition-
that pattern when reading non-words like ‘chead’, ing environments as opposed to other environments,
the conflict between the two patterns could, in part, e.g. how often <a> was read as /e/ before <nge> as
account for why people do not reliably read such opposed to <nce>.
words with the conditioned pronunciation, /ɛ/, as All the groups that we studied had a positive score
often as we would expect: other factors besides the on this measure: e.g. /e/ was used significantly more
coda /d/ pull them towards other pronunciations. often in the conditioning environment. This sensitiv-
Such a scenario seems very likely and definitely ity to context was found even among the first-graders
worthy of further investigation. that we studied. However, when we grouped the first-
graders by reading ability, we found that those read-
ing at a kindergarten level did not show sensitivity to
6 Course of Acquisition and context; indeed, most of their pronunciations did not
Use of Context in Reading match how the letters are pronounced in any English
word. Those reading at the first-grade level, however,
The previous studies all examined reading and did evince sensitivity to context.
writing behaviour in adults, or in computer As for the specific types of patterns, children read-
simulations that had been trained to adult levels. ing at first-grade level were sensitive to only three of

Writing Systems Research, Vol. 1, No. 1, 2009 27

B. Kessler

the patterns we investigated. They pronounced <a> patterns children are exposed to changes as they
as /ɑ/ more often after <w> and <qu> than after other progress through school. After all, the highly
onset letters; they pronounced <o> as /o/ more often context-dependent patterns we have been investi-
before <ld> and <lt> than before other coda letters gating mostly constitute what are often considered
in the absence of final <e>; and they pronounced irregularities in the orthographic system; perhaps
<oo> as /ʊ/ more often before <k> than before textbooks intentionally hold off on irregular words
other coda letters. At higher reading levels, children until children get older. An examination of the words
were sensitive to all eight patterns. The only excep- that appear in a graded corpus (Zeno et al., 1995)
tion was that the high-school students in our sample showed that this is not the case: The proportion of
showed an atypical reluctance to read <ea> as /ɛ/ such words in books remains remarkably constant
before <d>. from first grade on. Therefore, we cannot hypoth-
As for the magnitude of the effect, first-grade esize that first-graders are less likely to read <ea>
readers already showed an appreciable difference before <d> as /ɛ/ because they are not exposed to
of about fourteen percentage points between how so-called irregular words like ‘head’ and ‘dead’.
often they used a conditioned vowel in condition- Alternatively, perhaps children change their read-
ing versus non-conditioning contexts. As reading ing strategies as they get older or more mature,
grade improved, so did this magnitude. This change gradually becoming more and more convinced
was due both to increases in children’s produc- that context is a good thing. A simpler explana-

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

tion of the conditioned vowel where appropriate tion is that children get better at patterns as they get
(e.g. producing /ʊ/ for <oo> before <k>) and to more experience and practice with them. Given the
decreases in producing that vowel where not appro- complexity of English, it would not be surpris-
priate (e.g. for /ʊ/ <oo> before <n> or <m>). With ing that the course of improvement might take
the exception of a small dip at the high-school level five years or more. We tested this hypothesis by
due to that rejection of <ea> = /ɛ/, both trajecto- running an implementation of the connectionist
ries continued throughout our sample and into the reading model described by Harm and Seidenberg
college-age population described earlier. However, (2004). It was repeatedly exposed to 3,102 mono-
after about Grade 5, the improvements were small syllabic words, and its training was periodically
and not statistically significant. interrupted to test how it performed on our read-
Thus, children acquire and apply conditional ing task. Importantly, the vocabulary and read-
orthographic patterns from the very beginning of ing strategy of the simulation never changed, but
their careers as readers. Presumably these rather the trajectory of its output on the reading task as
obscure patterns are not taught overtly but are picked its exposure to vocabulary increased was very
up through general statistical learning, although it is, similar to the differences between children of dif-
of course, impossible to prove that children never ferent reading levels. It started off in early trials
have heard anyone else mention the rule. These with a profile similar to that of our first-graders, got
results do more damage to the rime-unit theory of progressively and rapidly better at discriminating
reading, because the children were at least as sensi- between the conditioning and non-conditioning
tive to patterns where the onset influences the vowel environments, then began improving much more
pronunciation (e.g. <a> as /ɑ/ after <w> or <qu>) gradually after about 500,000 epochs of training
as to those where the influence is the coda; indeed, (full details are at http://spell.psychology.wustl.
the onset pattern appears to be one of the first ones edu/InflCContxtOnPronV). While it is true that
acquired. Thus, children do not simply read rimes as no model perfectly emulates human behaviour,
whole units in an unconditional fashion. the simulation did show that a progression like
If children learn and apply contextual patterns, that of our grade-school students can be accounted
why does the magnitude of the effect remain below for entirely by gradual learning over time, using
adult levels for the first several grades of school? constant vocabulary and no change in learning styles
One reasonable hypothesis is that the vocabulary or reading strategy.

28 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

they observed the first effects of codas on vowels.

7 Course of Acquisition and Also as in the reading study, analysis of children’s
Use of Context in Writing texts show that the relevant vocabulary statistics
remain constant from first grade on, so differences in
Treiman and Kessler (2006) looked at children’s children’s textual environment cannot explain why
acquisition of conditional patterns from the other children’s spelling changes as they progress through
angle, that of spelling. We used a testing procedure grade school.
very similar to the one that had been used for adults: In other respects, the course of spelling develop-
testing how people would spell vowels for non-words ment was different from that of reading. Children
that have different types of consonantal contexts, as were slower to use context, with only very small
per Table 2. However, because spelling large num- effects in the kindergarten/first-grade group and
ber of words can tax the patience of young children, with acquisition not levelling off until the sixth-to-
we used a fill-in-the-blank procedure, asking them, eighth-grade group, as compared with fifth grade in
for example, to complete the spelling of the word the reading study. Although the sound-to-letter pat-
dictated as /glɛd/ by filling in what was missing on terns we chose for the spelling task were a little more
the answer form, which had <gl__d> already written. difficult than the letter-to-sound patterns used in the
Aside from making the task easier, this procedure reading task, as suggested by the Vocabulary statis-
also made the results somewhat cleaner, because the tics in Tables 2 and 3, it seems more likely that the

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

participant had no opportunity to select unintended nature of the two tasks is primarily responsible for
(and non-conditioning) consonants. One may object the differences. If children learn patterns primarily
that filling in a blank is an unnatural way to write by being exposed to them in print, then it is reason-
except in crossword puzzles and classroom exercises able to assume that access to the patterns would be
and so may draw on different skills than spelling activated more readily when seeing patterns already
a whole word, but this objection is largely countered printed, in the reading task. In writing, the stimuli are
by the observation that the oldest participants in this auditory, and young children not yet used to visual-
study, high-schoolers, behaved very similarly to the izing their spelling productions in advance may find
adults who had to spell entire words in the previous it hard to recognize that they are violating a spell-
study. The most important drawback was that one ing pattern until they have already written a word,
pattern from the adult study had to be omitted: the presumably with its most common, unconditional,
test of how /o/ is spelt before /l/. The basic choices vowel spelling.
are between <ole> and <oll>, so providing the coda The slow emergence of context sensitivity in
consonant or consonants in advance could have spelling affords us the opportunity to observe the
strongly biased the results either towards or away order in which different patterns emerge. Table 4
from using final <e>. lists for all eight patterns the spelling-grade level
For analysing the results, the participants were at which use of the pattern is first reliably attested.
grouped by spelling-grade levels (two or three A comparison with Table 2 shows that the age of
grades in a group) as determined by a standard first use does not correlate in any obvious way
spelling test (Jastak and Wilkinson, 1993). In many with how strongly the context conditions the vowel
ways the results were the same as for the child spelling (the Vocabulary statistic) nor with how
reading experiment. The kindergarten/first-grade often adults observe the context in their own spell-
group showed only a slight preference for using ings (Responses). However, Table 4 shows that the
conditioned vowels in conditioning contexts, but age of first use correlates with the frequency of the
the magnitude of the effect increased through grade most common unconditional spelling for the vowel
school. There was no evidence for rime-unit spell- in question. Although one must be very cautious
ing; in fact, children began observing a condition- about post hoc speculation on just eight data points,
ing effect of an onset on vowels (that /ɚ/ is mostly an enticing theory is that young children are more
<or> after /w/) at least three grade levels earlier than likely to pay attention to context when there is no

Writing Systems Research, Vol. 1, No. 1, 2009 29

B. Kessler

Table 4 Conditional vowel spelling patterns arranged by frequency of modal spelling

Vowel Modal correspondencea Conditional correspondence

Spelling Proportionb Pattern Grade attested

/ɚ/ <ir> 0.25 <or> after /w/ Kindergarten–1

/u/ <oo> 0.36 <oo> after non-coronals 2–3
/i/ <ea> 0.41 <ee> before /d/, /p/ 4–5
/ʊ/ <oo> 0.47 <oo> before /k/ 6–8
/ɑ/ <o> 0.51 <a> after /w/ 4–5
/aʊ/ <ou> 0.63 <ow> before /l/, /n/ 4–5
/aɪ/ <i> 0.66 <igh> before /t/ 6–8
/ɛ/ <e> 0.74 <ea> before /d/ 6–8
a
Most common spelling of the vowel across all contexts in monosyllabic English words.
b
Proportion of instances of the vowel that use the modal spelling.

candidate that clearly dominates across contexts. of these tasks, children as young as second grade
To put what must be a complex, implicit process were sensitive to some of these patterns, generat-

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

into crude terms, it is as if they quickly learn that ing or preferring spellings like <sull> and <sool>.
their best unconditional guess for spelling /ɚ/ will be The magnitude of the effect increased with each age
correct only 25% of the time, so they had better pay group considered. Magnitudes of the effect were
attention to conditional rules. In contrast, guessing also much greater in the recognition task than in the
<e> for /ɛ/ works 74% of the time, so the payoff for writing task, across all age groups. Finally, the effect
learning to use conditional rules is not very great. was much stronger for <(c)k> and <(l)l> than for
<(f)f> and <(t)ch>, in both tasks and across all age
groups; this point probably reflects the fact that the
8 Consonant Spelling general rule has some very high-frequency excep-
tions in the latter two cases, namely, ‘if’, ‘of’, ‘much’,
This article has focused up till now on vowels, and ‘such’.
primarily because of the disproportionately high The higher robustness of the effect in this study
number of important contexts that condition their as compared with the vowel spelling study may
sound–spelling correspondences. But some English be due to the fact that here, the conditional rule is
consonant spellings are conditioned by the vowel, entirely graphic. That is, <ll> would simply be ille-
and these have inspired some informative studies gal following <soo>, regardless of what sound one
as well. Hayes et al. (2006) looked at how children were spelling. The fact that these graphic rules are
and adults spell word-final consonants. As a first applied at younger grade levels than the condi-
approximation, this rule states that <f>, <l>, <k>, and tional sound-to-letter rules studied earlier suggests
<ch> are extended to <ff>, <ll>, <ck>, and <tch>, again that children find purely visual patterns easier
respectively, lexeme-finally after a stressed vowel to work with. The conclusion that there is a visual
if the rime would otherwise be spelt with just two advantage is strongly reinforced by the fact that
letters: <stiff>, <pill>, <back>, <patch>, not ∗<stif>, the same data show the effect much more strongly
∗<pil>, etc.; as opposed to <lift>, <life>, <loaf>, when presented in a purely visual task (‘which looks
and so forth, where the extension is not needed. We better, <sul> or <sull>?’) than when the participants
asked participants from the age of 7 to 22 years to had to generate spellings based on an auditory stimu-
spell pairs of non-words like /səl/ versus /sul/, or lus (‘spell /səl/’).
to choose which of two spellings like <sul> versus Hayes et al. (2006) also studied one conditional
<sull> or <sool> versus <sooll> looks better. In both rule affecting the spelling of onsets: that /k/ cannot

30 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

be spelt <c> but is normally spelt <k> before <e>, before front vowels in many words (‘get’, ‘gem’),
<i>, or <y>. We found that children as young as but it is perhaps surprising that our participants
second grade were more likely to spell /k/ as <k> pronounced <c> as /k/ a substantial number of times
in non-words before those letters than before <a>, when it appeared in non-words before <e> or <i>,
<o>, or <u>. a pattern virtually unknown in English. For exam-
ple, when asked to pronounce monosyllables like
<cersh>, they produced /k/ 16% of the time, but for
9 Consonant Reading <garsh>, nobody used /dʒ/. This imbalance appears
to be another instance of people’s tendency to never
One last question concerns the reading of conso- quite let go of the most common unconditional pro-
nants. Treiman et al. (2007) looked at how condi- nunciation of a letter, even when a specific context
tional rules affect the way adults read onsets. We does not support it. The same type of error has been
were interested in the fact that the pronunciations of reported in the Romance languages, even though
word-initial <c> and <g> are not only conditioned the rules for front and back pronunciations of <c>
by the following vowel, but also by the different and <g> apply in a very large number of words and
subsystems of English orthography. In the terminol- are virtually exceptionless (Content and Peereman,
ogy of Albrow (1972), in System 1, the basic system, 1992, for French; Job et al., 1998, for Italian).
word-initial prevocalic <g> is always /ɡ/, and <c> We also conducted several experiments where

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

is always /k/ but is used only before the historically the vocabulary stratum of the word was indicated
back vowel letters, <a>, <o>, and <u>. In System by graphic properties of the word’s spelling. In
2, the Romance system, both consonants have two one experiment, we constructed triplets such that
pronunciations, depending on whether they occur a nonsense syllable was presented either by itself
before the historically back vowels, where <g> is /ɡ/ (<geb>), or with a Romance suffix (<gebic>), or
and <c> is /k/, as in the basic system, or before the with a basic suffix (<gebful>). The stimulus <gebic>
historically front vowels, <e>, <i>, and <y>, where is overtly Romance, and so the pronunciation /dʒ/
<g> is /dʒ/ and <c> is /s/. (For brevity, I will hence- is expected, while <geb> and <gebful> have no
forth talk about front and back vowels and the front explicit markers for any spelling system (because
(/dʒ/ and /s/) and back (/ɡ/ and /k/) pronunciations basic suffixes are common even on Romance
of the consonants.) Thus, the pronunciation of these bases), and so could be pronounced either way. In
two letters is conditioned in English by the interac- another experiment, we constructed pairs of non-
tion of two separate factors: the identity of the fol- words that were as equivalent as possible, except
lowing vowel and the orthographic system to which that one member of the pair, such as <gireak>, used
the word or word component belongs. In a series of spelling patterns that are typical in the basic sys-
experiments, we asked college-age adults to read tem, and the other, such as <girec>, used typically
aloud a variety of words and non-words that begin Romance spelling patterns. Here, one would expect
with <c> or <g>, to see what conditions would influ- to hear more /ɡ/ in the former case and more /dʒ/ in
ence them to select front or back pronunciations for the latter case. We found that our participants did
those consonants. tend to use front pronunciations before front vowels,
In all the experiments, the participants showed the Romance pattern, more often when the word had
sensitivity to the identity of the following vowel. overt clues that it was Romance than when it did not.
Before back vowels, the consonants were almost This difference was significant, though it was small.
always given a back pronunciation, in accordance For example, <g> before <e> was pronounced /dʒ/
with an almost exceptionless rule that applies in 15% of the time in stimuli that had a basic suffix,
all systems. Before front vowels, both front and and 24% of the time in stimuli that had a Romance
back pronunciations were given for the consonant. suffix.
All factors being equal, this is not surprising for Our experiments showed that adult readers have
<g>, which does appear with both pronunciations learnt a fairly abstract conditional rule. Whereas most

Writing Systems Research, Vol. 1, No. 1, 2009 31

B. Kessler

rules discussed up to now have only required the place for increased attention to conditional patterns
reader or speller to consult the adjacent letter or in literacy instruction after all.
sound, this rule requires the reader to evaluate
a variety of clues to guess the appropriate spelling
system. Some of these clues involve looking ahead Funding
quite a few letters: sometimes only the end of a word This work was supported by the (US) National
tells how the initial letter should be pronounced. Institutes of Health [R01HD051610-02].
Lookahead is a challenge for models of reading such
as the DRC (Coltheart et al., 2001), which assume
that processing is strictly sequential.
References
Albrow, K. H. (1972). The English Writing System: Notes
10 Conclusion Toward a Description. London: Longman Publishing
Group.
Although English orthography appears chaotic, there Andrews, S. and Scarratt, D. R. (1998). Rule and
are many conditional rules that significantly increase analogy mechanisms in reading nonwords: hough
dou peapel rede gnew wirds? Journal of Experimental
its consistency. Such rules are somewhat compli-
Psychology: Human Perception and Performance, 24(4):
cated, but a large body of evidence shows that even 1052–86.

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

beginners process English in a way that cannot be
Aronoff, M. and Koch, E. (1996). Context-sensitive
accounted for by unconditional strategies such as
regularities in English vowel spelling. Reading and
whole-word or rime-unit memorization. Children Writing: An Interdisciplinary Journal, 8(3): 251–65.
take into account the identity of adjacent units when
Balota, D. A., Cortese, M. J., Sergent-Marshall, S. D.,
spelling consonants or vowels, and adults, at least,
Spieler, D. H., and Yap, M. J. (2004). Visual word
take into account relatively long-distance contexts, recognition of single-syllable words. Journal of
looking at the spelling or morphology of the rest of Experimental Psychology: General, 133(2): 283–316.
the word in order to decide how to pronounce the
Brown, G. D. A. (1998). The endpoint of skilled word
first letter. At present, we are investigating to what recognition: the ROAR model. In Metsala, J. L.
extent children can do this as well. and Ehri, L. C. (eds), Word Recognition in Beginning
Our main motivation in this work is to under- Literacy. Mahwah, NJ: Lawrence Erlbaum Associates,
stand the processes by which people learn and pro- pp. 121–38.
cess complex writing systems, but it is tempting to Caravolas, M. (2004). Spelling development in alphabetic
speculate briefly on paedagogical applications. The writing systems: a cross-linguistic perspective. European
obscurity of many of the patterns we have inves- Psychologist, 9(1): 3–14.
tigated, combined with the testimony of teachers Carney, E. (1994). A Survey of English Spelling. London:
(Hayes et al., 2006), strongly suggests that children Routledge.
are not explicitly taught these patterns but pick them Coltheart, M., Rastle, K., Perry, C., Langdon, R.,
on their own through general processes of statisti- and Ziegler, J. (2001). DRC: a dual route cascaded
cal learning: observing what letters, sounds, and model of visual word recognition and reading aloud.
sound–spelling correspondences are repeatedly Psychological Review, 108(1): 204–56.
found together in text. To a certain extent, there- Content, A. and Peereman, R. (1992). Single and
fore, these findings validate whole-language or multiple process models of print to speech
text-based approaches to literacy instruction. On the conversion. In Alegria, J., Holender, D., Morais, J.,
other hand, we have also shown that several useful and Radeau, M (eds), Analytic Approaches to Human
spelling patterns are not acquired until fairly late in Cognition. Amsterdam: Elsevier Science Publishers,
grade school, and in virtually all cases, the patterns pp. 213–36.
that are acquired are not applied nearly as often as Cook, V. (2004). The English Writing System. London:
they should be. This suggests that there may be a Arnold.

32 Writing Systems Research, Vol. 1, No. 1, 2009

 Conditional orthographic correspondences

Cummings, D. W. (1988). American English Spelling: Norris, D. (1994). A quantitative multiple-levels model
An Informal Description. Baltimore, MD: Johns Hopkins of reading aloud. Journal of Experimental Psychology:
University Press. Human Perception and Performance, 20, 1212–32.
Flexner, S. B. (ed.) (1987). Random House Dictionary, Nusbaum, H. C., Pisoni, D. B., and Davis, C. K. (1984).
2nd edn. New York: Random House. Sizing Up the Hoosier Mental Lexicon: Measuring the
Familiarity of 20,000 Words. Bloomington: Indiana
Glushko, R. J. (1979). The organization and synthesis University Press.
of orthographic knowledge in reading aloud. Journal
of Experimental Psychology: Human Perception and Plaut, D. C. and McClelland, J. L. (1993). Generalization
Performance, 5: 674–91. with componential attractors: word and nonword reading
in an attractor network. In Proceedings of the Fifteenth
Goswami, U. (1993). Phonological skills and learning to Annual Conference of the Cognitive Science Society.
read. In Tallal, P., Galaburda, A. M., Llinas, R. R., and Hillsdale, NJ: Lawrence Erlbaum Associates, pp. 824–9.
von Euler, C. (eds), Temporal Information Processing
Plaut, D. C., McClelland, J. L., Seidenberg, M. S.,
in the Nervous System: Special Reference to Dyslexia
and Patterson, K. E. (1996). Understanding normal
and Dysphasia. New York: New York Academy of
and impaired word reading: computational principles
Sciences, pp. 296–311.
in quasi-regular domains. Psychological Review, 103:
Hanna, P. R., Hanna, J. S., Hodges, R. E., 56–115.
and Rudorf, E. H. (1966). Phoneme–grapheme Powell, D., Plaut, D. C., and Funnell, E. (2001).
Correspondences as Cues to Spelling Improvement. A Developmental Evaluation of the Plaut, McClelland,

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

Washington, DC: U.S. Department of Health, Seidenberg, & Patterson (1996) Connectionist
Education, and Welfare. Model of Single Word Reading [Abstract]. Worcester,
Harm, M. W. and Seidenberg, M. S. (2004). Computing UK: Meeting of the British Psychological Society,
the meanings of words in reading: cooperative division Developmental and Education Sections.
of labor between phonological and visual processes. Rumelhart, D. E. and McClelland, J. L. (1986).
Psychological Review, 111(3): 662–720. Parallel Distributed Processing: Explorations in the
Hayes, H., Treiman, R., and Kessler, B. (2006). Children Microstructure of Cognition. Cambridge, MA: MIT
use vowels to help them spell consonants. Journal of Press.
Experimental Child Psychology, 94(1): 27–42. Ryder, R. J. and Pearson, P. D. (1980). Influence
of type-token frequencies and final consonants on
International Phonetic Association (1999). Handbook of
adults’ internalization of vowel digraphs. Journal of
the International Phonetic Association: A Guide to the
Educational Psychology, 72(5): 618–24.
Use of the International Phonetic Alphabet. Cambridge,
England: Cambridge University Press. Stanback, M. L. (1992). Syllable and rime patterns
for teaching reading: analysis of a frequency-based
Jastak, S. and Wilkinson, G. S. (1993). Wide Range vocabulary of 17,602 words. Annals of Dyslexia, 42:
Achievement Test, 3rd edn. Wilmington, DE: Jastak 196–221.
Associates.
Treiman, R. and Kessler, B. (1995). In defense of an
Job, R., Peressotti, F., and Cusinato, A. (1998). onset–rhyme syllable structure for English. Language
Lexical effects in naming pseudowords in shal- and Speech, 38(2): 127–42.
low orthographies: further empirical data. Journal of Treiman, R. and Kessler, B. (2006). Spelling as statistical
Experimental Psychology: Human Perception & learning: using consonantal context to spell vowels.
Performance, 24(2): 622–30. Journal of Educational Psychology, 98(3): 642–52.
Johnson, D. D. and Venezky, R. L. (1976). Models Treiman, R., Kessler, B., and Bick, S. (2002). Context
for predicting how adults pronounce vowel digraph sensitivity in the spelling of English vowels. Journal of
spellings in unfamiliar words. Visible Language, 10(3): Memory and Language, 47(3): 448–68.
257–68.
Treiman, R., Kessler, B., and Bick, S. (2003). Influence
Kessler, B. and Treiman, R. (2001). Relationships of consonantal context on the pronunciation of vowels:
between sounds and letters in English monosyllables. a comparison of human readers and computational
Journal of Memory and Language, 44(4): 592–617. models. Cognition, 88(1): 49–78.

Writing Systems Research, Vol. 1, No. 1, 2009 33

B. Kessler

Treiman, R., Kessler, B., and Evans, R. (2007). Antici- the International Phonetic Association (1999), and are
patory conditioning of spelling-to-sound translation. enclosed in slash marks. The Appendix gives keywords
Journal of Memory and Language, 56(2): 229–45. illustrating how I applied the symbols of the IPA to
Treiman, R., Kessler, B., Zevin, J. D., Bick, S., and American English.
Davis, M. (2006). Influence of consonantal context
on the reading of vowels: evidence from children.
Journal of Experimental Child Psychology, 93(1): 1–24. Appendix
Treiman, R., Mullennix, J., Bijeljac-Babic, R., and
Table A1 Keywords for use of International Phonetic
Richmond-Welty, E. D. (1995). The special role of
rimes in the description, use, and acquisition of English Alphabet symbols for American English
orthography. Journal of Experimental Psychology: IPA Keyword
General, 124: 107–36.
aɪ ice
Treiman, R. and Zukowki, A. (1988). Units in reading
aʊ out
and spelling. Journal of Memory and Language, 27:
æ add
466–77. ɑ odd
Venezky, R. L. (1970). The Structure of English d dig
Orthography. The Hague: Mouton. dʒ jump
ð this
Venezky, R. L. (1999). The American Way of Spelling: e ace

Downloaded from wsr.oxfordjournals.org by guest on May 26, 2011

The Structure and Origins of American English ε ebb
Orthography. New York: Guilford. ə up
Wilkinson, G. S. (1993). Wide Range Achievement Test 3. ɚ earn
Wilmington, DE: Wide Range, Inc. f fine
ɡ good
Zeno, S. M., Ivenz, S. H., Millard, R. T., and Duvvuri, R.
i eat
(1995). Educator’s Word Frequency Guide. Brewster, k kiss
NY: Touchstone Applied Science Associates. l log
Ziegler, J. C., Stone, G. O., and Jacobs, A. M. (1997). m mouse
What is the pronunciation for -ough and the spelling n nice
for /u/?: a database for computing feedforward and ŋ sing
feedback consistency in English. Behavior Research o oak
Methods, Instruments & Computers, 29: 600–18. ɔ ought
p pie
Zorzi, M., Houghton, G., and Butterworth, B. (1998). ɹ red
Two routes or one in reading aloud?: a connectionist dual s sun
process model. Journal of Experimental Psychology: ʃ ship
Human Perception and Performance, 24(4): 1131–61. t toy
tʃ cheese
u ooze
ʊ book
Notes v voice
1 Graphemes of the orthography are enclosed in angled w wind
θ think
brackets. Phonemes are represented by symbols of

34 Writing Systems Research, Vol. 1, No. 1, 2009