You are viewing the SNL 2018 Archive Website. For the latest information, see the Current Website.

Poster D69, Friday, August 17, 4:45 – 6:30 pm, Room 2000AB

Discern: a computational model of naming deficits in bilingual speakers with aphasia

Claudia Penaloza1, Uli Grasemann2, Maria Dekhtyar1, Risto Miikkulainen2, Swathi Kiran1;1Boston University, 2The University of Texas at Austin

Several factors including the combination of spoken languages, the relative competency in the two languages, and the effect of brain damage in bilingual adults with aphasia (BAA) make it challenging to examine bilingual naming impairment and rehabilitation clinically without large scale longitudinal studies. DISCERN is a computational model that allows to systematically examine lexical access in healthy bilinguals and BAA, thus offering a potential solution to simulate treatment response and predict optimal rehabilitation outcomes in BAA. The DISCERN model is a neural network simulation of lexical access that consists of three interconnected self-organizing maps: one semantic and two phonological maps for L1 and L2 respectively. All maps are linked by adaptive associative connections and get trained to encode the semantic and phonetic representations of words in both languages (i.e., 651 unique words in English and corresponding translations in Spanish). We used an evolutionary algorithm (Bäck, 1996) to identify the best-fit training schedule (i.e., set of training parameters including number of words trained per simulated year, learning rates and neighborhood size, age, age of acquisition (AoA) and language exposure) that would allow training each individual neural network to effectively simulate each healthy participant’s naming performance. Twenty-one healthy Spanish-English bilinguals, one Spanish and one English monolingual were recruited for simulations of healthy naming performance (i.e., Boston Naming test-BNT and a 60-item naming screener) and 24 Spanish-English BAA (34.86 ± 46.61 months post-stroke) for simulations of language impairment (i.e., Pyramid and Palm Trees- PAPT and BNT). The best-fit training schedule allowed simulating naming in the healthy participants, explaining 78% of the variance in their naming scores. This best-fit training schedule was used together with the premorbid data of the BAA (i.e., age, AoA and language exposure) to generate pre-stroke naming models. These models were then lesioned systematically to simulate damage to the bilingual lexical system. Semantic impairment was first modeled by applying noise to the semantic map at different intensities until it matched each patient’s post-stroke semantic performance (i.e., PAPT scores). Damage to the semantic map impacted naming in each language (i.e., BNT performance) reflecting naming impairment associated with semantic deficits. Additional damage to the phonetic maps was also able to account for naming deficits in addition to semantic impairment. Post-lesion models with damage to the semantic and phonetic maps accurately matched the post-stroke PAPT scores and lead to a reasonable simulation of bilingual naming impairment for each BAA. These results show how naming impairment of patients with different language impairment profiles can be accurately modeled by applying lesion damage to their individual neural network models. Such individual models of naming impairment in BAA will be used for treatment simulations in future research. Retraining lesioned models in one or the other language can help understand the influence of pre-morbid language exposure, AoA, and other factors on bilingual aphasia treatment outcomes.

Topic Area: Computational Approaches