Attention-Based Aspect-Level Sentiment Mining for Data-Driven Market Research: An Explainable AI Perspective on Consumer Feedback Analytics

Aspect-based sentiment analysis captures opinions on specific product and service features, but accurate and interpretable aspect-level prediction remains difficult for texts with multiple aspects and overlapping opinions. This paper presents XAT-Miner, an explainable framework that combines an aspect-aware attention module, VADER and TF-IDF features, an ensemble classification head, attention-based token highlights and an aspect-wise market sentiment index. With static pretrained word vectors (92,532 trainable parameters, CPU training in under a minute), XAT-Miner reaches 71.04% accuracy and a macro F1-score of 0.532 on 908 held-out SemEval-2014 Task 4 aspect instances, on par with TF-IDF with a linear SVM (70.93%, 0.529). Cross-validation, McNemar tests and an incremental ablation show that the lexical features carry most of the signal. In the static-vector model, attention highlights influence the prediction 2.5 to 4.8 times more than random tokens, and the market index computed from out-of-fold predictions follows the gold label index (Spearman ρ=0.98). On a fine-tuned RoBERTa-base backbone, the same design reaches 78.5% accuracy, matching plain fine-tuning (78.0%), although its attention weights are weak highlights and leave-one-out scores are preferable. The framework thus offers a lightweight CPU option and an accurate contextual option for consumer feedback analytics.