A Neuro-Computational Approach to Understanding the Mental Lexicon

Human lexical knowledge does not appear to be organised to minimise storage, but rather to maximise processing efficiency. The way lexical information is stored reflects the way it is dynamically processed, accessed and retrieved. A detailed analysis of the way words are memorised, of the dynamic interaction between lexical representations and distribution and degrees of regularity in input data, can shed some light on the emergence of structures and relations within fully-stored words. We believe that a bottom-up investigation of low-level memory and processing functions can help understand the cognitive mechanisms that govern word processing in the mental lexicon. Neuro-computational models can play an important role in this inquiry, as they help understand the dynamic nature of lexical representations by establishing an explanatory connection between lexical structures and processing models dictated by the micro-functions of human brain. Starting from some linguistic, psycholinguistic and neuro-physiological evidence supporting a dynamic view of the mental lexicon as an integrative system, we illustrate Temporal Self Organising-Maps (TSOMs), artificial neural networks that can model such a view by memorising time series of symbolic units (words) as routinized patterns of short-term node activation. On the basis of a simple pool of principles of adaptive Hebbian synchronisation, TSOMs can perceive possible surface relations between word forms and store them by partially overlapping activation patterns, reflecting gradient levels of lexical specificity, from holistic to decompositional lexical representations. We believe that TSOMs offer an algorithmic model of the emergence of high-level, global and language-specific morphological structure through the working of low-level, language-aspecific processing functions, thus promising to bridge the persisting gap between high-level principles of grammar architecture (lexicon vs. rules), computational correlates (storage vs. processing) and low-level principles and localisations of brain functions. Extensions of the current TSOM architecture are envisaged and their theoretical implications are discussed.