In text classification tasks with complex models and high-stakes domains the alignment between predictions and explanations tends to be weak because post-hoc explainability methods operate independent of model training. In this paper, we suggest ATM-AM - an approach based on the Gated Recurrent Unit (GRU) that combines Bahdanau attention with a training-time SHAP-backed alignment objective to offer real-time, context-aware interpretability without trade-off in predictive performance. The model is tested over three frequently-used sentiment analysis datasets (IMDbhttps://huggingface.co/datasets/imdb, Amazon Reviews https://www.kaggle.com/datasets/bittlingmayer/amazonreviews, and SST-2. https://huggingface.co/datasets/glue/viewer/sst2) yielding accuracy scores of 91.8%, 89.5%, and 90.0% with respective F1-scores of 0.899, 0.877, and 0.889 respectively, on each dataset. We also average our measurements over three runs for statistical soundness. The additional training latency added by ATM-AM is quite modest (13–18%), and the inference time remains short (3–4 ms per sample), rendering it feasible to be deployed in real-time. A user-centered interpretability study with 30 participants obtained an average rating of 4.6/5 showing that users trust the explanations produced by our proposed model. These observations posit ATM-AM as a feasible and interpretable solution Text Classification in contexts where model behavior needs to be accountable and reliable.
Pittala et al. (2026) studied this question.