A Cross-Linguistic Analysis of Aphasic Speech

Poster Session D - Monday, April 15, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Sreekar Baddepudi¹, Josh Van Zak²; ¹Evergreen Valley High School, San Jose, California, USA, ²Cambridge University

Wernicke's aphasia (WA) is a neurological disorder caused by damage, often from a stroke, to specific language networks in the human brain. This damage results in speech that is spoken at a normal speed and rhythm, and with proper grammatical structure, but the content is usually incomprehensible. Research on people with WA typically focuses on linguistic errors in relation to a person’s native language. Studying these errors has failed to yield clear insights regarding specific neural mechanisms and damage behind particular WA symptoms. For instance, if certain linguistic errors are shared among WA patients who speak languages with different structures and syntax, then this realization may point to specific neural networks that are affected compared to those that may be more language-specific. The analog here is low-level vs. high-level programming languages. To evaluate this set of hypotheses, transcripts from native speakers of English, French, Japanese, and Cantonese (WA and healthy controls) were collected from the AphasiaBank (https://aphasia.talkbank.org/). A state-of-the-art large language model (Anthropic’s Claude v2) was used to perform an in-depth comparison of speech components and rules, ranging from grammar to coherence, across different representative languages. The model was able to identify signs of Wernicke’s Aphasia, such as neologisms, paraphasias, filler words, and word finding difficulty, as well as many commonalities between the languages. Based on this data, an interlinguistic computational analysis model for aphasic speech has been developed.

Topic Area: LANGUAGE: Other