Kessler, Brett & Rebecca Treiman. 2004, June. Sensitivity to statistical contextual patterns when spelling consonants in English. Paper presented at the meeting of the Society for the Scientific Study of Reading (SSSR), Amsterdam, Netherlands.


Several consonant phonemes in English can be spelled multiple ways, but often the correct spellings become more predictable when the vowel is taken into account. We measured 11 such patterns in the English lexicon, then tested the extent to which spellers make use of them. College students spelled pairs of monosyllabic nonwords with matching consonants but differing vowels. In all cases, spellers were sensitive to vowels when choosing consonant letters. The computational model of Houghton & Zorzi showed similar tendencies, but was not sensitive to all the patterns humans were sensitive to.


Spelling in English is challenging because almost all phonemes can be spelled more than one way, e.g., /s/ is <s>, <se>, <c>, <ce>, etc. To lessen the need for rote memorization, spellers learn and sometimes exploit the frequencies of sound-letter correspondences (Barry & Seymour, 1988). The odds of choosing the correct letters could be further improved if spellers learned how these frequency distributions may be modified by phonetic context—e.g., while /s/ is usually <se> after /aʊ/ as in "house", <ce> is more likely after /aɪ/ as in "nice". Treiman, Kessler, and Bick (2002) documented that spellers exploit context when spelling vowels, but the question is still moot for consonants.

Experiment 1 tested whether people (26 Detroit college students) exploit the effect of the vowel on the onset distributions. 3 onsets were studied: /k/, /sk/, and /g/. For each onset, 10 pairs of monosyllabic nonwords were presented that differed only by vowel, e.g., /kaɪp/ ~ /kaʊp/. Vowel pairs were selected such that the first vowel conditions a distinctive spelling of the onset—here <k>—more often than the second vowel; this was confirmed by an analysis of all familiar English words. Experiment 2 was analogous, but with a separate group of 24 students spelling 8 different codas (/dʒ/, /tʃ/, /f/, /k/, /l/, /s/, /z/, /ɡ/). We tested, for example, whether students would use the distinctive spelling <ce> for /s/ more often after vowels like /aɪ/ than vowels like /aʊ/. The nonwords were presented in random order, with 20 fillers in each experiment. In all cases, participants used the distinctive spelling significantly (p < .05) more often before the conditioning vowel, whether analysed by subjects or by items (nonword pairs).

This evidence for contextual spelling argues against simple models such as phoneme-by-phoneme context free spelling or spelling by whole onsets and rimes; the latter would not predict an influence of the vowel on the onset. More promising are implicitly statistical spelling models such as the two-layer neural net proposed by Houghton & Zorzi (2003). We built an implementation using General American spellings and pronunciations, and trained it on vocabulary familiar to US college students (Kessler & Treiman, 2001). We used the implementation to test the same nonwords as in the experiments above. For all 11 consonants, the net agreed with the human participants in activating the distinctive spelling more often when spelling consonants next to conditioning vowels. However, the strength and consistency of the effect was much weaker than with humans. For example, the sound /k/ activated the distinctive spelling <k> more often before the conditioning vowel /ɪ/, but not before other vowels humans were also sensitive to (/i/, /ɛ/, and /aɪ/). Although the net models human spelling well in many respects, a more complex architecture will be required in order to fully account for human performance.


APA citation:

Kessler, B. & Treiman, R. (2004, June). Sensitivity to statistical contextual patterns when spelling consonants in English. Paper presented at the meeting of the Society for the Scientific Study of Reading, Amsterdam, Netherlands.

Last change 2009-08-07T11:09:46-0500