Lexical decision is a common psycholinguistic task. It is quite simple in principle. Participants see some letters or hear some sound, and they must respond with whether what they saw or heard is a word. For this post, we will focus on visual lexical decision. There are quite a few effects that provide insights into how humans process languages. For example, people are more likely to make errors on nonwords like CHARE. This is because reading CHARE aloud usually sounds like CHAIR. Even during reading, auditory processing is also being used. Errors are also more likely on nonwords like TIKCET. This suggests humans are not reading letters sequentially, and that order matters less than most people believe.
Baboons and machines have been shown to be able to complete this task as well, after some training . While learning TensorFlow, I implemented a computational version of this task that takes images of potential words and determines their lexicality using a convolutional neural network. In the model, success is based on learning the regularities in words (e.g. XK is not found in any word while EE is quite common) and on rote memorization of frequent strings. Drawbacks include the lack of an auditory component, making it impossible to simulate effects like (pseudo)homophony. Nevertheless, it provides a simple example of visual categorization using convnets and is a starting point for computational models of reading that make use of visual, auditory, and semantic representations that are grounded in the real world.
Featured image from . My code with comments can be found in this github repository.
References Hannagan, Thomas, Johannes C. Ziegler, Stéphane Dufau, Joël Fagot, and Jonathan Grainger. “Deep learning of orthographic representations in baboons.” PloS one 9, no. 1 (2014): e84843. doi:10.1371/journal.pone.0084843  Herzig, Daniela & Sullivan, Sarah & Evans, Jonathan & Corcoran, Rhiannon & Mohr, Christine. (2012). Hemispheric asymmetry and theory of mind: Is there an association? Cognitive Neuropsychiatry. 17. 371-96. doi:10.1080/13546805.2011.643556.