What can we learn from learning models about sensitivity to letter-order in visual word recognition?

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psycholinguistics

Background:

• Previous research indicates high reader tolerance for letter transpositions in Indo-European languages.
• This led to computational reading models emphasizing universal flexibility in positional coding.
• However, these models fail to account for cross-linguistic variations in transposed-letter effects.

Purpose of the Study:

• To investigate how linguistic environments shape transposed-letter effects in computational models.
• To challenge the notion of universal flexibility in reading positional coding.
• To explore the role of orthographic properties in visual word recognition.

Main Methods:

• A domain-general connectionist architecture was trained on words from diverse linguistic environments.
• The model performed tasks involving letter transposition and substitution.
• Analysis focused on how learned orthographic statistics influenced performance.

Main Results:

• Model performance demonstrated that letter-order coding flexibility is influenced by language-specific orthographic properties.
• The prevalence of anagrams within a language significantly impacts transposed-letter effects.
• The model's learning process revealed insights into neurobiological noise in letter-position registration.

Conclusions:

• Reading flexibility is not solely a universal principle but is shaped by statistical orthographic properties of a language.
• Computational learning models offer a powerful tool for understanding visual word recognition.
• Novel predictions for empirical research were generated, highlighting the advantage of learning-based approaches.