What is so hard about abstract words, anyways? A neuromechanistic explanation using brain-constrained neural network models

It is well known that children first learn words with concrete meaning, whereas abstract ones are typically acquired later. To obtain clues about the underlying causes and mechanisms, a brain-constrained neural network model of brain regions involved in semantic processing was trained on instances of concrete and abstract concepts. Model areas contained excitatory and inhibitory neurons, with Hebbian and anti-Hebbian synaptic plasticity rules. This allowed for the formation of neuronal circuits representing different concepts. Concept formation was simulated based on the similarity structure of objects, actions and scenes exemplifying instances of the concept. Subsequently, conceptual instances were co-presented with wordforms to establish semantic links and direct grounding of symbols in the world. Instances of concrete concepts (e.g. different DOGs or HAMMERs) lead to conceptual representations in the form of reverberant neuronal circuits. Such concept formation was impossible for abstract concepts (e.g. BEAUTY, FORCE): the resultant neuronal circuits failed to reverberate and only responded selectively to their grounding instances. When learning wordforms for the different concepts, abstract words developed fully functional neural circuits, comparable with those of concrete words. However, the formation of these circuits was delayed. While concrete concepts were learned preverbally and could then be ‘labelled’ by a word, abstract concepts only emerged when learned with a wordform. This suggests that the delay in abstract word acquisition might be caused by structural dissimilarities between the real-world events these words are used to speak about.