Poster D4, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Neurocomputational model of the mental lexicon in a word naming and word retrieval scenario
Catharina Marie Stille1, Trevor Bekolay2,3, Stefan Heim4,5,6, Bernd J. Kröger1;1Department for Phoniatrics, Pedaudiology and Communication Disorders, Medical Faculty RWTH Aachen University, 2Applied Brain Research, Waterloo, Canada, 3Centre for Theoretical Neuroscience, University of Waterloo, 4Department of Psychiatry, Psychotherapy, and Psychosomatics, Medical Faculty, RWTH Aachen University, 5Research Centre Jülich, Institute of Neuroscience and Medicine (INM-1), Germany, 6JARA — Translational Brain Medicine, Aachen
The relationship between neural dysfunctions from deficits in specific brain regions and behavioral deficits like lexical disorders is an ongoing research topic. Models of speech production and speech perception offer an approach for relating neural deficits to behavioral deficits. The aim of the present study is to simulate a naming task with semantic and phonological cues under the condition of defined neural deficits. The simulated model is a large-scale neural model of speech production. This model is implemented with spiking neurons (leaky-integrate-and-fire neurons; LIF neurons) using the NEF (Neural Engineering Framework) and the SPA (Semantic Pointer Architecture) (Eliasmith et al., 2012; Eliasmith, 2013). The model comprises a cognitive processing module, a three-level knowledge repository (with concept, lemma and lexeme levels; the mental lexicon) and a cortico-cortical control loop for action selection that includes the basal ganglia and thalamus (Stewart, Choo & Eliasmith 2010 a, b). In total, the model comprises 19 neuron buffers with 3250 neurons per buffer (61750 LIF neurons in total). Each neuron ensemble within the basal ganglia and thalamus comprise 50 neurons, yielding 2100 neurons in the basal ganglia and 400 neurons in thalamus. Learned items in the semantic and lexeme levels are organized as semantic pointer networks (cf. Kröger et al., 2016). Here, our conceptual network comprises 874 surface concepts and 222 deep concepts (e.g., superordinates like “vegetables” and abstract items like “love”). The lexeme network comprises 889 surface forms representing syllables or words and 1194 deep forms representing subsyllabic structures like single speech sounds and sound clusters (i.e., consonant clusters). The relationship between surface and deep forms models phonological and semantic relationships between surface items in a specific language. Neural deficits are modeled by decreasing neural activity in different buffers of the mental lexicon, i.e. within specific buffers that represent concepts, lemmas or lexemes in the perception or production speech processing pathways. Speech behaviors, i.e., naming deficits that occur for a specific neural lesion, are simulated here by using the model to perform a concrete picture naming task used in logopedic diagnostics comprising semantic and phonological cues (WWT 6-10; Word Range and Word Retrieval Test; Glück, 2011). The main question of our simulation experiments is: are phonological or semantic cues helpful in a naming task if neural deficits occur in specific cortical locations dealing with lexical processing and storage? Our simulation results indicate that (i) semantic and phonological cues facilitate naming even if neural dysfunctions occur within cortical locations dealing with speech processing and lexical storage; (ii) phonological cues have a stronger facilitating effect for naming than semantic cues. The only case in which semantic cues are more effective than phonological cues is if a neural deficit is located strictly in the concept storage part of the mental lexicon.
Themes: Computational Approaches, Meaning: Lexical Semantics
Method: Computational Modeling