Modeling public opinion emotional structure in emergencies through cross-conditional analysis of the SMP2020-EWECT Weibo benchmark

In emergencies, social media is one of the main platforms on which people share how they feel, and those expressions carry information that matters for situational awareness and risk communication. This study models public opinion through fine-grained emotion classification on a public benchmark, the SMP2020-EWECT Weibo dataset, which is labeled with six emotions: happy, angry, sad, fear, surprise and neutral. In the emergency subset the happy label rarely marks private cheerfulness. It marks encouragement, mutual support and praise directed at medical staff and at the affected city, a crisis-specific reading of the label that the token-level analysis in Sect. 6 confirms, and for that reason this class is referred to throughout as the positive, solidarity-oriented class. The benchmark contains two subsets that share the same label scheme and the same platform. The usual subset reflects ordinary online discussion, while the virus subset was collected during the COVID-19 outbreak and represents the emergency condition. Comparing the two subsets is the main analytical device of the study, because it shows how the emotional structure of public expression differs between an emergency and an ordinary period. A bidirectional gated recurrent unit encoder followed by an additive attention layer is trained separately on each subset and is compared with strong linear baselines built on term frequency features, with a variant of the same encoder in which attention is replaced by mean pooling, and with a Transformer encoder of matched scale that shares every other component of the pipeline. Emoji tokens and hashtag topic words are kept during preprocessing because they carry sentiment on microblogs, and a class-weighted loss counters the strong imbalance of the rare emotions. All neural results are reported as the mean and standard deviation over five random seeds. The attention model reaches an accuracy of 0.679 ± 0.005 on the general subset and 0.685 ± 0.011 on the emergency subset. The linear baselines are higher, at 0.716 and 0.734; an exact McNemar test shows that this advantage is statistically significant on both subsets, and a Transformer encoder trained under the same protocol does not close the gap. The attention layer therefore contributes interpretability rather than accuracy, and a token deletion test quantifies that contribution: deleting the three tokens that receive the highest attention in a post costs 16.8 accuracy points on the emergency subset and 15.2 points on the general subset, against 6.9 and 3.4 points for deleting three tokens at random. The emotional structure itself differs sharply between the two conditions, with the positive, solidarity-oriented class covering 51.4% of the emergency subset, or 48.7% once repeated posts are removed from the training split, while anger covers 30.0% of the general subset, a contrast that is directly informative for monitoring. Because the benchmark records neither who wrote a post nor which posts were removed, the study cannot separate spontaneous solidarity from expression shaped by state-led mobilization or platform moderation, and the solidarity-oriented share is interpreted as the composition of visible posts. The result is a transparent and reproducible account of how the structure of visible public emotion shifts during an emergency.