Identifying Customer Priorities in Online Reviews through Sequence-to-Sequence Learning with Dual Contextual Attention

User reviews on e-commerce platforms offer real-time feedback on customer experiences. However, the large volume and unstructured nature of reviews make it challenging for businesses to identify the key aspects that matter most to consumers. This study introduces a sequence-to-sequence learning model that integrates global positional attention with local syntactic attention to extract aspect terms capturing the product and service attributes emphasized in reviews. A multi-stage aspect summarization process, combining dimensionality reduction, semantic clustering, and LLM based refinement, is utilized to distill hundreds of extracted terms into a concise set of interpretable primal words highlighting customer major concerns. Experiments on three annotated SemEval 2014 datasets demonstrate superior extraction accuracy compared to eight established baselines. Application to five large Amazon review datasets across different domains further shows strong generalizability in identifying coherent primal terms for each dataset, providing actionable themes aligned with consumer priorities.