Developing an online hate classifier for multiple social media platforms

Table 12 Relative differences of linguistic variables between comments predicted as hateful by XGBoost + All and those labeled as hateful in the ground truth

LIWC category	Rel. diff. (lower scores) (%)	LIWC category	Rel diff. (higher scores) (%)
Parenth^a	− 13.4	Friend^a	+ 6.9
Quote^a	− 8.6	Body^a	+ 5.3
Dash^a	− 7.8	Swear^a	+ 5.2
QMark	− 5.6	Sexual^a	+ 5.1
WC	− 4.9	Bio	+ 4.7
Risk	− 4.7	Informal	+ 4.6
anx	− 4.5	Anger	+ 4.4
Work	− 4.4	Semic	+ 4.0
Tone	− 4.0	Netspeak	+ 3.6

Relative difference is calculated as C_predicted − C_ground)/C_ground, where C is a LIWC category
^aOutlier: > 1.5 interquartile ranges (IQRs) below the first quartile or above the third quartile