Inside Semantic Feature Analysis: A Within-Trial Analysis of Feature Quantity and Quality

Purpose: Aphasia rehabilitation increasingly emphasizes the importance of understanding the mechanisms and ingredients underlying behavioral interventions. Semantic feature analysis (SFA) is a commonly used intervention for anomia in aphasia, but there is little evidence directly examining its active ingredients. Using within-trial responses during SFA, we sought to explore how generating semantic features might facilitate naming improvements. Method: A retrospective analysis evaluated data collected from a clinical trial focused on intensive SFA treatment for individuals with chronic aphasia. The study included 44 adults with chronic aphasia following left-hemisphere stroke. A pretrained semantic model was used to estimate the semantic relatedness between features and targets. Results: Participants were 2.6 (95% confidence interval CI; 2.16, 3.12) times more likely to correctly name target words after engaging in feature generation. Each additional feature generated was associated with a three percentage-point increase (95% confidence interval CI; 0.2, 0.4) in the probability of a correct response at the end of the trial; a 1 SD increase in semantic similarity (i.e., feature quality) was associated with a six percentage-point increase (95% CI 0.4, 0.10). Model comparison favored semantic similarity over feature generation count in predicting final response accuracy. Conclusions: Findings provide converging evidence that semantic feature generation is an active ingredient in SFA treatment, emphasizing the importance of feature quantity and semantic quality, consistent with a spreading activation account of SFA's benefits. Further research is warranted to validate relatedness values from semantic model embeddings and to explore the relationship between within-trial feature generation and generalization to semantically related but untreated words. Supplemental Material: https://doi.org/10.23641/asha.32159406