Clinical Aphasiology Conference 2023 Poster Presentation
Automating intended target identification for paraphasias in discourse using a large language model
We predicted the intended target words of paraphasias in discourse using a machine learning based large language model (LLM). Data consisted of Cinderella story retelling transcripts from people with aphasia. Human transcribers determined the intended target of the paraphasias. We replaced paraphasias with a blank token and fine-tuned the LLM to fill in the blank with a predicted target, based on the rest of the story retelling. The model achieved 46.8% accuracy. It performed significantly better on targets with perfect human agreement and higher human confidence in target identification, and on paraphasias from participants with less severe or fluent aphasia.