Context and prior phonological knowledge as a support for novel word learning: A computational study with the...

According to the "universal" theory of self-teaching, phonological decoding acts as a self-teaching mechanism, allowing word-specific orthographic knowledge to be acquired incidentally during reading. Empirical evidence suggests that word-specific context and prior phonological knowledge support self-teaching across languages, especially for inconsistent words and in less advanced readers. The computational models of self-teaching developed so far are dual-route models. They predict a major role for decoding in orthographic learning and postulate the involvement of context for phonologically known, partially decoded words. In this work, we describe BRAID-Acq, a single-route computational model of self-teaching. In three simulations, we tested the model's ability to generate phonological forms of novel words (consistent or inconsistent) using only lexical knowledge. We further assessed learning outcomes across four conditions defined by the presence or absence of prior phonological knowledge and/or contextual information. Simulation results show that the BRAID-Acq model successfully acquires new orthographic and phonological representations through reading, autonomously detecting whether words were novel or familiar, even without context. Learning benefited from two corrective mechanisms: phonological lexical feedback and contextual pronunciation correction, which allowed the correction of transient mispronunciations for orally known words without generating errors for unknown words. Simulations showed that context and phonological knowledge had the greatest effect for inconsistent words and for intermediate-sized lexicons. A further simulation showed that context was robust, supporting learning across a wide range of sizes and strengths without causing context-induced errors. The overall findings provide a proof-of-concept that self-teaching can be successfully implemented in a single-route framework.