Contribution
The paper gives fake and true Weibo news a five-emotion profile (anger, disgust, joy, sadness, fear) from a hand-labelled lexicon, and finds that anger separates fake from true news and highly from lowly retweeted news, with joy the other way. It adds regression models with controls, English datasets from Twitter and news media, a survey of reposting motives, and a proposed intervention: flag posts whose anger share passes 20%.
Key insight
Instead of explaining fake news’s spread by its diffusion structure, which is observed only after the fact, the authors look at what is in the post at the moment it is posted. Among the five emotions they score, only anger and joy differ consistently across every grouping and every scoring method, in opposite directions.
Assumptions
- Labels. Fake news is what Weibo’s committee of outside fact-checkers tagged false; it is tagged “after being widely disseminated” (S1). True news is untagged news from verified users (media, elites, authorities). So the two sets differ by selection as well as truth: the fake set is enriched for widely spread posts, the true set for verified authors.
- Emotion is the fraction of emotional terms. A post’s share of anger is its anger terms over its emotional terms. The lexicon has 6,155 terms (1,323 anger, 710 disgust, 2,066 joy, 1,243 sadness, 813 fear) from 34,227 candidates, each labelled three times by nine coders aged 18–30, keeping terms on which the labels agreed. No agreement statistic or validation against human-rated posts is reported. It covers 87.1% of posts; the rest count as neutral.
- Groups. Posts are split by author followers and retweets. The text gives the optimal split as (x*, y*) = (10, 1000), but S2.2–S2.3 and S10 use 10³ followers and 10 retweets, so the tuple is printed in the wrong order. “Highly retweeted” therefore means at least 10 retweets.
- The retweet outcome is the count within 48 hours of posting (S9).
- Age is not controlled, on the argument that most Weibo users are 18–30 and that topic absorbs the rest (S8).
Key results
- Distributions (Fig. 1, Tables S2–S5). Mean anger share: HLT 0.111 against LHF 0.267 (K-S D = 0.275); true 0.112 against fake 0.165 (D = 0.101); L 0.123 against H 0.240 (D = 0.210). Mean joy: HLT 0.611 against LHF 0.329 (D = 0.366); true 0.609 against fake 0.443. Fear is higher in fake news (0.091 against 0.222). Disgust shows no consistent difference.
- Robustness of the distributions (S7). A naive-Bayes classifier (accuracy 0.64), a neural classifier (0.69) and a rank measure over six embedding classifiers (0.67–0.86) reproduce the anger and joy gaps. The other negative emotions flip between methods, and are thereafter folded into “other emotions”.
- Regressions (Table 1). Lexicon shares, with follower, friend,
mention, hashtag, location, date, URL, length, emergency and topic
controls, robust standard errors.
- Logit, fake against true: anger 0.889 (s.e. 0.097) and joy −1.507 for HF against LT news (N = 10,486); anger 0.385 and joy −1.279 for all news (N = 26,831). Pseudo-R² 0.35–0.36.
- Linear, retweets in 48 h: anger 23.96 and joy −29.56 in fake news (model 3, N = 20,323, R² = 0.084); anger 22.28 and joy −35.98 in all news (model 4, N = 26,831, R² = 0.134).
- The emotion shares sum to 1, so each coefficient is relative to the omitted bundle of disgust, sadness and fear.
- The headline number. “Increasing the occupation of anger by 0.1 and reducing that of joy by 0.1 in fake news leads to 5.8 more retweets.” From model 3 (fake news) that is 2.40 + 2.96 = 5.4. The 5.8 is model 4 (all news): 2.23 + 3.60 = 5.83. Likewise “2.2 more retweets” for anger alone is model 4’s 0.1 × 22.28. Both numbers come from the all-news model and are attributed in the text to fake news.
- Other controls (Table 1, S8). The emergency dummy is one event, the 2015 Tianjin explosion: 663 fake posts and 3 true. Its logit coefficient of 5.6 reflects that near-separation, and in the retweet models it is negative (−33.5, −23.0). Mention is positive for retweets but not significant in model 1.
- English data (S10, Table S18). On 129,690 unlabelled COVID-19 and 2016-election tweets, scored with the NRC lexicon, anger’s logit coefficient for ≥ 10 retweets is 0.215 (s.e. 0.062) and joy’s −0.225; disgust, surprise and anticipation are also negative. No fake/true labels are involved. On the Kaggle fake/real news set, anger share is 0.123 against 0.110 (D = 0.108). On 478 PHEME rumour tweets, HF 0.142 against LF 0.020.
- COVID-19 Weibo case (S7.3). 200 lexicon-scored fake posts: anger minus joy is higher in HF than LF (t = 2.851, P = 0.006). No true news and no regression.
- Survey (Fig. 2, S11–S13). 15 posts (5 each from HLT, LHF, HLF), chosen near K-means cluster centres without regard to emotion; six questionnaires; 1,291 valid responses from 1,316 Weibo users aged 18–30, via wjx.cn. Four rumour-sharing motives on a four-point scale (Cronbach’s α 0.56–0.80), each centred on the respondent’s mean of the four. LHF posts draw more anxiety management than HLF and HLT (D = 0.235 and 0.144), and fake posts more information sharing than true. Grouping by dominant emotion, only two posts are anger-dominated (LHF News1 and News5) against five joy-dominated: anger draws more anxiety management (0.027 against −0.057, D = 0.134, p = 0.023) and information sharing (0.444 against 0.335, p = 0.042), and joy more self-enhancement. No correction for the many comparisons.
- Intervention (S14). Flagging posts with anger share above θ = 0.2 maximizes the gap between flagged HF and HT news: it would flag 46% of highly retweeted fake news and 22% of highly retweeted true news, and 89% of flagged H posts are fake. A logit gives an odds ratio of 2.651 for H-Anger, reported as “165% more likely to be fake”; that is odds, not probability. The effect assumes, “very optimistically”, that no flagged post is retweeted.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | On Weibo 2011–2016, fake news carries a larger share of anger words and a smaller share of joy words than true news | moderate | lexicon and three model families, Tables S2–S14; label selection unaddressed |
| C2 | Within the same data, anger share goes with more retweets in 48 h and joy share with fewer, after content, author and topic controls | moderate | Table 1 (3)–(4), Table S17; observational, R² ≤ 0.135 |
| C3 | Anger causes the spread of fake news | weak | the regressions in C2, read causally; no experiment or identification strategy |
| C4 | The anger–retweet association holds on Twitter regardless of platform and culture | weak | one logit on unlabelled tweets, Table S18; small fake/true gaps in S5 and S6 |
| C5 | Anger-dominated news draws stronger anxiety-management and information-sharing motives | weak | two anger-dominated stimuli, K-S tests at p ≈ 0.02–0.04 |
| C6 | Flagging posts above 20% anger would stop 46% of highly retweeted fake news at a cost of 22% of popular true news | weak | in-sample threshold search, with the assumption that flagged posts are never retweeted |
Concepts
- Occupation of an emotion. The fraction of a post’s emotional terms that carry it.
- LHF, HLF, HLT, … Fake (F) or true (T) news by author followers (L/H, split at 10³) and retweets (L/H, split at 10).
- Structural virality. Mean shortest-path distance over a diffusion tree (Goel et al.); about 2 is pure broadcast.
- Information dominance. ID(G_i, G_j), the normalized difference in transmission coverage between follower bands, used to justify the 10³ split.
- Anxiety management, information sharing, relationship management, self-enhancement. The four rumour-sharing motives of Sudhir & Unnithan’s scale.
Connections
It starts from Vosoughi, Roy & Aral (2018) on Twitter, whose findings on diffusion structure and speed it reproduces on Weibo (S2.3, S3), and from Berger & Milkman (2012) on arousing emotion and sharing, and Brady et al. (2017) on moral emotion and diffusion. The record’s other reading on information operations, LIT-083, is about sharing disinformation incidents among defenders, not about spread.
Bearing on the record
- No THEORY, QUESTION or CLAIM cites this paper. The reading supports one new account, filed as THEORY-329, held at the strength of an association.
- The LIT’s triage summary is right in outline, with three
corrections.
- “Raising anger by 0.1 and cutting joy by 0.1 adds about 5.8 retweets” is the all-news model; for fake news alone the model gives about 5.4.
- “Questionnaires tie anger-dominated news to anxiety-management and information-sharing motives” rests on two anger-dominated posts.
- “In regressions controlling for … emergencies” should say one emergency, the Tianjin explosion.
- Tags.
social-science,emotion-and-affectandsociety-and-governancefit;network-sciencewould fit the diffusion-structure supplement but is not where the claim lives. - No instruction for machine-learning practice. The tagging proposal is platform policy, not ML practice, and the classifiers are instruments.
Limitations
- Observational. The paper uses “causally” throughout for coefficients from logit and OLS models on observational data. There is no experiment, instrument or natural experiment. Unmeasured content features correlated with anger words (topic within the seven classes, rumour type, calls to “repost”) could carry the effect.
- Label selection. Fake news is labelled because it spread, and true news comes from verified accounts; the fake-against-true logits mix truth with this selection.
- Weak fit. The retweet models explain 8–13% of the variance.
- Measurement. The lexicon is unvalidated against post-level human judgement, and an emotion share is undefined for 13% of posts.
- Survey. Few stimuli (two anger-dominated), many unadjusted tests, α as low as 0.56, and motives measured as stated, not as reposting.
- Inconsistencies. The split tuple is printed reversed (S2.1); group sizes differ between Table S1 (LHF 1,397), Table S2 (1,238) and Table S16 (1,326) without explanation; the 5.8 is attributed to the wrong model.
Open questions
- Does manipulating a post’s anger, holding its claim fixed, change how often it is shared? A randomized exposure experiment would settle what the regressions cannot.
- Does flagging angry posts reduce sharing in practice? The threshold is chosen in-sample and its effect assumed, not measured.