Mapping Communication Disruption in Traumatic Brain Injury with Transformer Embeddings

Recent advances in computational modeling have expanded our capacity to analyze language and communication, particularly through transformer models. The present work investigates how such computational frameworks can be leveraged to address clinical domains in communication disorders. We used semantic embeddings from BERT's layers to analyze language-related adjustments used by participants with traumatic brain injury (TBI) in conversational transcripts. By examining semantic convergence patterns across different layers of the BERT model, we found that TBI participants demonstrated more pronounced "self" convergence -- they tended to stay closer to their own semantic contributions in the conversation -- compared to controls. This effect was particularly noticeable at earlier layers of the BERT model, suggesting that surface-level semantics play a significant role. The findings highlight the potential for language models to enhance our understanding of social interaction dynamics. We further discuss how bridging computational linguistics with clinical domains can address analytic challenges in the study of natural cognition and communication.