Application of integrated gradients explainability to sociopsychological semantic markers

The paper presents an application of the integrated gradients (IG) method for enhancing explainability in the classification of sociopsychological semantic markers, particularly focusing on agency using the BERTAgent model. It outlines a novel training procedure that promotes overfitting to improve class distinctiveness, even with limited labeled data, and evaluates the contributions of specific words to classification outcomes. This work is significant for practitioners as it provides a framework for interpreting model decisions in nuanced sentiment analysis, potentially improving the transparency and applicability of AI in social psychology contexts.