When a public emergency strikes, the emotional texture of social media changes, and a new study of Chinese microblogs has now measured that change with unusual candor. Working from the SMP2020-EWECT benchmark, a public collection of Weibo posts labeled with six emotions, researchers compared a subset gathered during the COVID-19 outbreak with a subset of ordinary online discussion. In the general subset, anger was the largest class, covering 30.0 percent of posts. In the emergency subset, the happy label dominated at 51.4 percent of posts, or 48.7 percent once repeated posts were removed from the training split. Yet the study is explicit that in the emergency condition the happy label rarely marks private cheerfulness. It marks encouragement, mutual support and praise directed at medical staff and at the affected city, which is why the authors call it the positive, solidarity-oriented class throughout.
The contrast between the two subsets is the central empirical finding, and the researchers are careful about what it does and does not establish. The two subsets share the same six labels, namely happy, angry, sad, fear, surprise and neutral, and the same platform, so the annotation scheme and data source are held fixed while the condition changes. The authors describe this as a paired subset comparison rather than a controlled experiment, because the subsets also differ in size, collection period and topical composition. The general training split contains 27,768 posts while the emergency training split contains 8,606, and the virus subset is drawn from a single event. The emotional contrast, they note, is a property of the labeled data that can be read directly from the label proportions; no deep model is needed to establish it, and none is claimed to do so.
What the models add is a transparent account of how such a structure can be detected and inspected. The team built a bidirectional gated recurrent unit encoder followed by an additive attention layer, trained separately on each subset, and compared it against strong linear baselines built on term frequency and inverse document frequency features, a variant in which attention was replaced by mean pooling, and a Transformer encoder of matched scale. Emoji tokens and hashtag topic words were kept during preprocessing because they carry sentiment on microblogs, and a class-weighted loss countered the strong imbalance of the rare emotions. Every neural result was reported as the mean and standard deviation over five random seeds, and all models were trained and evaluated separately on each subset so that the emergency and general conditions were never mixed.
The accuracy figures favor the simplest tools. The attention model reached an accuracy of 0.679 plus or minus 0.005 on the general subset and 0.685 plus or minus 0.011 on the emergency subset. The linear baselines were higher, at 0.716 and 0.734, and an exact McNemar test on the paired per-instance predictions showed that this advantage was statistically significant on both subsets. The linear accuracy on each subset lay above the full five-seed range of the attention model, which spanned 0.672 to 0.686 on the general subset and 0.666 to 0.700 on the emergency subset. A Transformer encoder trained under the identical protocol finished 1.3 accuracy points below the recurrent encoder with attention on the general subset and 2.5 points below it on the emergency subset, and 7.3 and 6.9 points respectively below the stronger linear baseline of each condition. The authors attribute this to the data: posts are short, with a median of twenty tokens or fewer, and the training splits are too small to train a Transformer from random initialization to its usual advantage.
The attention layer therefore contributes interpretability rather than accuracy, and the study quantified that contribution with a token deletion test. Deleting the three tokens that receive the highest attention in a post cost 16.8 accuracy points on the emergency subset and 15.2 points on the general subset, changing the predicted label of 38 percent and 44 percent of posts respectively. The tokens that receive high attention are thus the tokens the prediction depends on. The test also set a limit: deleting the three tokens ranked highest by the global term frequency coefficient for the predicted class cost 20.0 and 21.5 points, more damaging than the attention ranking on both subsets, so a corpus-level lexicon identifies decisive tokens at least as well as per-post attention in aggregate.
What attention provides that a global coefficient cannot is a weight that belongs to one post rather than to the corpus. When each weight was reported as a ratio to the uniform weight, the most attended token in a post received on average 11.0 times the uniform weight in the emergency subset. The encouragement token jiayou carried a mean ratio of 5.6 across the 828 emergency test posts in which it occurs, but its tenth and ninetieth percentiles were 0.56 and 12.7, a spread of more than an order of magnitude; the salute token had a mean ratio of 3.8 with percentiles of 0.23 and 9.1 over 537 posts. The rank correlation between the attention ordering and the global coefficient ordering was 0.19 on the emergency subset and 0.21 on the general subset, so the two orderings are related but far from interchangeable. Attention explains the prediction the deployed system actually made, token by token, which matters when an analyst must decide whether an individual alarm is trustworthy.
The study also confronted an uncomfortable property of its own data: repetition. The cleaned text of 237 of the 5,000 general test posts, or 4.7 percent, also occurred in the general training split, and the same held for 473 of the 3,000 emergency test posts, or 15.8 percent. In the emergency subset these matches were concentrated in short formulaic solidarity phrases, such as a city encouragement slogan or a single heart emoji, and in forwarded copies of widely shared posts. A deduplicated re-evaluation removed every such test post and re-scored all five systems without retraining. Every system lost accuracy, and the loss was several times larger in the emergency condition, where the overlap was larger: between 3.2 and 4.2 accuracy points and between 2.4 and 2.9 macro F1 points, against about one accuracy point on the general subset. The ordering of the systems was unchanged, and the McNemar test remained significant, with p equal to 3.5 times ten to the minus seventh on the emergency subset and 4.2 times ten to the minus twenty-seventh on the general subset.
Repetition also inflated the headline share of the solidarity class. Of the 8,606 emergency training posts, 1,044, or 12.1 percent, were additional copies of a string occurring elsewhere in the same split, and 71.7 percent of those copies carried the happy label. After exact deduplication the positive class fell from 51.4 to 48.7 percent of 7,549 distinct emergency strings, and to 47.6 percent under a broader normalized matching; anger in the general subset fell from 30.0 to 28.8 percent. The repetition itself was informative. Of the 24 emergency strings occurring ten or more times, 19 were labeled happy in all or most copies, and 14 were shorter than fifteen characters, mostly slogans such as Wuhan jiayou, which alone occurred 44 times. Among the 326 longer posts repeated five or more times, 78 percent carried the Weibo repost marker, two thirds of these forwards reproduced posts by just five entertainment accounts, and every one of those forwards carried the happy label. Forwarding was common throughout the emergency training split, where 39.8 percent of posts carried the repost marker, against 0.1 percent of the general training split.
These patterns bear directly on what the label proportions measure. The authors set out two compatible readings. Crisis and emergency risk communication theory describes how, during the acute phase of a public health crisis, official messaging, calls for collective action and norms of mutual aid encourage prosocial and efficacy-oriented expression, and the token evidence, dominated by appeals to a city and to medical staff, blessings for safety and collective pronouns, fits that account. A second explanation is equally compatible: solidarity was also promoted from above through state-led campaigns that popularized slogans such as Wuhan jiayou, moderation selectively removes posts with collective action potential, and government-affiliated posters have been shown to publish large volumes of cheerleading content. The benchmark records neither who wrote a post nor which posts were removed before collection, and it contains no timestamps or author identifiers, so the data cannot separate spontaneous solidarity from expression shaped by mobilization or platform moderation. The solidarity-oriented share is therefore interpreted as the composition of visible posts.
The study’s practical upshot is deliberately modest. On this benchmark and these two conditions, a transparent linear model is a strong and economical choice for emotion monitoring, with the neural model reserved for the per-post explanations that a corpus-level weighting cannot provide. Strong pretrained Chinese systems reach accuracy in the high seventies on this six-class task, such as 0.783 accuracy and 0.785 F1 for an ERNIE-based multi-channel model, and the results here sit roughly five to ten accuracy points below that level, a gap the authors read as the cost of a lightweight design rather than a competitive result. The rare emotions remain weakly modeled, with surprise supported by only 68 test posts in the emergency subset, and the benchmark covers one platform, one language and one health emergency, so generalization to other platforms, languages and event types is untested. What the work delivers is a reproducible account of how the structure of visible public emotion shifts during an emergency, and a demonstration that honest baselines and behavioral tests can say more about crisis expression than a leaderboard number alone.
Subject of Research: Fine-grained emotion classification of Weibo posts comparing an emergency COVID-19 subset with a general subset of the SMP2020-EWECT benchmark
Article Title: Modeling public opinion emotional structure in emergencies through cross-conditional analysis of the SMP2020-EWECT Weibo benchmark
Article References: Liu, X., & Qiao, X. (2026). Modeling public opinion emotional structure in emergencies through cross-conditional analysis of the SMP2020-EWECT Weibo benchmark. Discover Computing, 29(1), Article 847. https://doi.org/10.1007/s10791-026-10787-6
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10787-6
Keywords: Weibo, SMP2020-EWECT, emotion classification, COVID-19, public opinion, attention mechanism, solidarity, class imbalance, token deletion test, data duplication, risk communication, natural language processing
News Source: Denise Maddox. (November 4, 2026). During COVID-19, Weibo anger gave way to solidarity slogans, dataset reveals. Scienmag.
