January 6th on Twitter: measuring social media attitudes towards the Capitol riot through unhealthy online conversation and sentiment analysis

Figures & data

Figure 1. The study methodology we used.

Table 1. Mean ROC-AUC score of the classifiers.

Baseline (Price et al., 2020)	LSTM (Gilda et al., 2022)	LR	RF	SVM	BERT	RoBERTa	XLNet
0.7446	0.7121	0.7068	0.6802	0.6602	0.7745	0.8026	0.8075

Figure 2. ROC curves and ROC-AUC scores for each attribute, for the XLNet model we trained.

Figure 3. Tweet trends by unhealthy attribute cluster and by the proportion of unhealthy tweets relative to the total number of tweets.

Figure 4. Correlogram of unhealthy attributes vs. the NRCLex sentiments, computed with Cramer’s V-measure.

Figure 5. Radar plots showing the proportion of tweets exhibiting the sentiment for each sub-attribute.

Table 2. Most relevant topics.

Attribute	Name	Topic
antagonize	anti_trump	gop, traitor, republican, president, party, america, american, election, loser, impeachment
antagonize, hostile	accusation_racism	white, capitol, maga, terrorist, insurrection, supporter, racist, black, woman, house, white, capitol, terrorist, maga, racist, american, black, supporter, house, crime
condescending	treason	resident, gop, traitor, election, republican, party, america, vote, country, american
generalization	anti_voting	democrat, republican, american, president, America, left, country, party, never, vote
generalization	far_right_white_supremacist	white, supporter, maga, terrorist, capitol, racist, right, supremacist, nazi, boy
generalization	proud_boys	white, black, supporter, racist, right, woman, supremacist, maga, boy, proud
generalization	right_or_republican	republican, gop, party, democrat, american, vote, dems, election, voter, left
generalization	left_or_liberal	american, democrat, america, left, hate, black, woman, country, liberal, president
hostile	trump_coward	gop, traitor, republican, party, president, election, america, vote, country, coward
sarcastic	sarcastic_keywords	yeah, lol, right, job, president, thing, make, hell, better, well, wow, gop, week, hey, rudy, maybe, twitter, big, check, riot
healthy	capitol_riot	election, republican, right, vote, capitol, president, party, even, insurrection, state
healthy	election_or_impeachment	president, first, order, house, administration, executive, covid, american, impeachment, senate
–	vulgar_abuse	f*ck, sh*t, stupid, f*cking, care, hell, dumb, shut, maga, back

Figure 6. Network formed of LDA topics (nodes) linked by cosine similarity scores (edges). Only scores ≥ 0.5 have been included.

Table 3. Topics computed for attribute-emotion pairs (selection).

Attribute-emotion pair	Topic
dismissive_anger_1	traitor, lies, lie, stop, lying, fraud, one, people, delusional
hostile_trust_0	president, people, gop, us, god, one, america, never, country
generalization_fear_0	republicans, us, right, terrorists, maga, white, left, party, gop
generalization_fear_1	people, hate, democrats, supporters, america, police, violence, maga, gop
generalization_trust_0	republicans, people, supporters, gop, president, want, americans, democrats, left
generalization_trust_1	white, supporters, people, right, maga, police, supremacists, one, black
sarcastic_fear_0	god, would, oh, hell, us, bad, yeah, know, gop
sarcastic_fear_1	maga, people, one, need, let, us, police, 2, time
sarcastic_anger_0	sh*t, people, maga, us, would, one, think, wow, stop
sarcastic_anger_1	hell, gop, bad, know, damn, oh, house, let, hey

Price, I., Gifford-Moore, J., Flemming, J., Musker, S., Roichman, M., Sylvain, G., Thain, N., Dixon, L., & Sorensen, J. (2020). Six attributes of unhealthy conversations. Proceedings of the Fourth Workshop on Online Abuse and Harms, 114–124. https://doi.org/10.18653/v1/2020.alw-1.15

 Google Scholar

Gilda, S., Giovanini, L., Silva, M., & Oliveira, D. (2022). Predicting different types of subtle toxicity in unhealthy online conversations. Procedia Computer Science, 198, 360–366. https://doi.org/10.1016/j.procs.2021.12.254

 Google Scholar