An information-theoretic analysis of the inflectional regular-irregular gradient for optimal processing units

Prediction-driven word processing defines the human ability to anticipate upcoming input words in recognition. From this perspective, input word forms need to be processed as quickly and efficiently as possible. Under the reasonable assumption that spoken words are memorized and processed as word trees (e.g. Marslen-Wilson's "cohorts"), the larger the size of the cohort of an input word at a certain point in time (and the later its uniqueness point), the harder and slower to process the word is. Regularly and irregularly inflected verb forms have different stem family sizes and different uniqueness points. Using a Recurrent Neural Network (RNN) as a computational model of the human lexical proces- sor, we explore here how their distributional and structural properties may affect (optimal) processing strategies.