<p>Extracting user preferences from reviews to enhance recommender systems is a prevalent research focus. However, existing methods face two critical challenges. First, when leveraging auxiliary (stranger) reviews to alleviate data sparsity, they often lack an effective mechanism to select those reviews that are truly pertinent to the target item, and thus treat all auxiliary reviews as equally informative. Second, in modeling fine-grained aspects, traditional topic models are context-agnostic, whereas randomly initialized aspect vectors in neural methods lack semantic coherence, both failing to capture latent topics accurately. To address these issues, we propose BEACON, a BERTopic-Augmented Context-Aware Co-Attention Recommender System. We initially develop a Context-Aware Auxiliary Review Selection Module that leverages the target item’s global context to collaboratively attend to and select key user and auxiliary reviews, enabling the model to focus on the most relevant textual evidence. We then introduce a BERTopic-Enhanced Aspect-Guided Representation Module, which leverages the powerful contextual modeling capabilities of BERTopic to discover high-quality, semantically coherent aspects automatically and utilizes them to learn user and item representations. On top of these components, a Factorization Machine Rating Prediction Module captures higher-order feature interactions among ID and aspect-aware features for final rating prediction. Comprehensive experiments on multiple Amazon datasets demonstrate the effectiveness and superiority of BEACON, which consistently surpasses strong baselines in rating prediction accuracy.</p>