How does orthography affect the way word sounds are represented in our minds?

Paper: Althaus, Kotzor, Schuster & Lahiri (2022). Cognition

Code (Eye tracking data preprocessing in Python and modelling in R): Github

Figure1. Eye movement patterns for three types of verbs in the fragment completion paradigm (Shaded areas reflect standard error of the mean).

The problem

In many languages there are words like “two” vs. “too” and “toe” vs. “tow” that sound the same but differ in spelling. How does this affect our mental representations of them? Similarly there are words that sound different even though their spelling is the same, e.g. “read” (present tense, [ri:d]) vs. “read” (past tense, [rɛd]). The difficulty in examining these mental representations, and how phonology and orthography interact, is that such cases are not numerous.

The study

In Bengali verb inflection there is a completely regular vowel raising process that allows us to investigate this systematically. In Bengali the first and third person singular either have the same or a raised stem vowel. For instance, ‘mare’ (s/he hits) becomes ‘mari’ (I hit; no stem vowel raising, merely the suffix changes to indicate 1st Person), but khæle ‘he/she plays’ changes to kheli ‘I play’ (i.e. here the [æ] is raised to [e] in the first person), and similarly for lekhe/likhi (s/he writes / I write) the stem vowel is also raised. Most importantly, only the e-i alternation (but not the æ-e alternation) is reflected in a spelling change (i.e. ae and e in khaele, kheli have the same Bengali spelling).
This gave us a way to test how speakers of Bengali represent these sounds in the mental lexicon with a "fragment completion task": we played them the first syllable of a verb (e.g. khæ- or khe-) and their task was to determine whether this came from khæle or from kheli. For their response, they saw both options written in Bengali letters on a screen and were asked to press a button corresponding to the target – but we also recorded their eye movements with an eye tracker.


We expected participants to be unable to decide when hearing “ma”, because “mari” and “mare” share the stem vowel. That’s exactly what we found. By contrast, we expected participants to find this easy whenever the stem vowel changes (either from [æ] to [e], or from [e] to [i]). Any differences between pairs like khæle/kheli (no spelling change) and lekhi/likhi (spelling change) should then be attributed to the impact of orthography, because the differences in vowel are perceptually similar in both cases. However, it turns out that participants were really good at performing the task for the lekhe/likhi group of words (with a spelling difference), but surprisingly bad at finding the correct answer for the khaele/khele-type words (where there is no difference in spelling). In fact, their manual responses for these words were at chance – only their eye movements showed that in fact they could solve the question. Apparently spelling has a really big impact on how we “hear” sounds.
We determined this by examining the time course of looking from the first target the partcipant looked at. In general, if the target is seen first, participants tend to stay there, whereas if the distracter is seen first, they move away to the other item – but when we compare trials on which the subject started at target vs. distracter, the distance between those data plots tells us how easy they found it to decide: If the plots are close together, they could not decide (target/distracter lead to the same responses), whereas if they are far apart, they found it easy (they looked at the target more than the distracter).

Modelling Approach

Statistically, to determine whether the eye movement patterns were distinct, we used a growth curve approach (Mirman, 2014). This involves fitting polynomials to the time course data (using lme4 in R and logistic models) and successively added fixed effects to the base model. Model comparisons then showed that all three word types were different from each other; for the mare-mari group of words there was no difference between target-first and distracter-first (i.e. subjects had to guess), but both khæle-kheli and lekhe-lekhi word groups were discriminated. It’s also clear, though, that the participants found the decision much easier for lekhe-lekhi (large distance, even early on) than for khæle-kheli (small distance throughout).


What the results overall mean is that orthography affects our mental representations of phonological patterns. We believe that exposure to the differences in spelling over and over again (and the need to produce the correct spelling in writing) means that the mental representation of those word forms become more distinct. Where there is no spelling change, the mental representations for 1st and 3rd person forms overlap to a large extent, which makes them confusable. In this experiment, once the mental representation is activated, memory for the acoustic input decays rapidly, and this may mean that both forms seem like good competitors for those items where the mental representation overlaps.


Mirman, D. (2014). Growth curve analysis and visualization using R. Boca Raton, FL: CRC Press / Taylor & Francis Group.