Table 4.

Best-performing methods within each text mining method family

	Feedback data				Election data
Family	Method	F-score	Acc.	Subset acc.	Method	F-score	Acc.	Subset acc.
RMD	Dict. 1	.329	.800	.152	Dict. 1	.492	.875	.312
CMD	Dict. 2	.537	.806	.260	Dict. 3	.477	.879	.298
SML	RoBERTa	.779	.919	.562	RoBERTa	.696	.939	.550
Zero-shot	GPT-4	.499	.666	.032	GPT-4	.409	.836	.227
	Reddit data				Hate speech dataa
Family	Method	F-score	Acc.	Subset acc.	Method	F-score	Acc.	Subset acc.
RMD	Dict. 1	.369	.734	.195	Dict. 1	.611	.611	NA
CMD	Dict. 3	.501	.837	.376	Dict. 3	.845	.845	NA
SML	RoBERTa	.690	.927	.626	RoBERTa	.907	.907	NA
Zero-shot	GPT-3.5	.492	.861	.386	GPT-4	.897	.897	NA

All performances are reported on the holdout dataset, which is identical for each method

^a In the Hate speech dataset, different categories cannot co-occur in a single text, which is why micro-averaged score and accuracy take the same values and the subset accuracy is not calculated