xAITC: an approach for explaining and improving opaque-based depression classification

Deep Learning models have been widely used for effective text classification in domains ranging from medicine to computer networks. Unfortunately, explaining how the opaque models operate is challenging due to inherent complexities. Explainable Artificial Intelligence has been utilized to address this issue. Nevertheless, approaches that use explainability to understand how opaque models make decisions often fall short in using their results to improve the explained model. This paper introduces xAITC, an approach based on explainability techniques to understand and enhance, in a post-hoc manner, text classification models that follow deep neural network architectures, particularly encoder-only transformers. xAITC utilizes global and local explainability to identify the least relevant words to remove from the datasets for further re-training of the models, seeking model enhancement. xAITC was showcased to explain and improve an encoder-only transformer model used for depression classification in a COVID-19 Twitter dataset. Results revealed that xAITC is a promising solution for improving text classification-centered models based on deep neural networks, as enforcing xAITC improved all measured learning metrics, namely accuracy, balanced accuracy, and F1-score, in the encoder-only model. Remarkably, these improvements led to better detection of non-depressive and depressive tweets.