Abstract: Product safety incidents on social media often evolve rapidly into complex and multidimensional public opinion crises. These crises are reflected not only in rising online attention but also in fine-grained semantic signals, including risk perceptions, public concerns, responsibility attribution, official responses, emotional intensity, and collective mobilization. Existing public opinion monitoring methods primarily rely on coarse sentiment polarity, keyword frequency, or simple propagation indicators, which limits their ability to capture the semantic complexity and dynamic evolution of product safety crises. In addition, conventional multi-label text classification methods generally treat labels as independent categories, which makes it difficult to model hierarchical relations and co-occurrence constraints among crisis-related labels. As a result, both the predictive accuracy and the interpretability of public opinion risk analysis remain limited. To address these challenges, this study uses Weibo data from a Romoss power bank safety incident as an empirical case and constructs a manually annotated multidimensional semantic dataset for product safety public opinion analysis. Building on this foundation, this study proposes a deep learning model, HPLR-Net, which integrates hierarchy-aware prompt modeling, label-query representation, and lightweight label re-ranking. This architecture enables the accurate identification of key semantic categories while preserving consistency among related labels. The model focuses on a set of learnable primary labels, including Risk, Doubt, Mobilize, Official, and Corp.doubt, and further incorporates emotional intensity and attitude recognition to provide a more comprehensive characterization of public opinion dynamics. To connect post-level semantic recognition with temporal risk analysis, this study further introduces an out-of-fold semantic monitoring mechanism. The predicted probabilities of the main semantic labels are aggregated by date and combined with raw propagation statistics, including posting volume and comments, to construct a composite public opinion risk index (CPORI). The index follows a semantic-priority weighting scheme in which risk descriptions, public doubts, and emotional intensity are given primary importance, while posting volume and comments are incorporated as indicators or propagation intensity following logarithmic transformation. An analytic hierarchy process and weight sensitivity analysis are used to examine the interpretability and robustness of the weighting scheme. Experimental results show that HPLR-Net outperforms several baseline models in the main-label semantic recognition task and has particular advantages for difficult labels and label-structure consistency. The CPORI series effectively captures the outbreak, escalation, and decline stages of the crisis, and its high-risk days show good temporal alignment with external event anchors. The results further indicate that the decline of public attention does not necessarily mean the simultaneous restoration of risk perception or public trust because Risk, Doubt, and negative attitude signals may remain elevated after posting volume has decreased. To examine the transferability of the proposed workflow, the trained HPLR-Net and the same CPORI construction procedure are also applied to Weibo posts related to the Samsung Note7 battery explosion incident. The external-event application shows that the method can generate an interpretable daily risk curve for another consumer electronics safety incident without retraining the model. Overall, this study provides an integrated framework for fine-grained semantic recognition, composite risk index construction, and dynamic trend interpretation in product safety public opinion monitoring. The findings offer practical support for regulatory early warning, enterprise crisis management, and public safety risk governance in social media environments.