Table 2.

Results of the three best feature configurations in the LOOCV setting and on the test set in terms of accuracy.

Feature configurations	Input modality	LOOCV accuracy	Test set accuracy
Temporal + char4grams	audio + text	0.8611	0.9167
New + char4grams	audio + text	0.8889	0.8750
char4grams	text	0.8611	0.8958
top three late fusion	/	0.8796	0.9375
BERT—reimplementation of Yuan et al. (2020)	/	0.8426	0.8333
ERNIE best related work (Yuan et al., 2020)	/	/	0.8958

The feature configurations column indicates which feature configuration has been used and whether char4grams have been added, and column Input modality shows the modality on which ADR features have been generated. The best individual methods' results in LOOCV and on the test set, as well as the late fusion of all three methods, are shown in bold. The row labelled top three late fusion presents the results of employing late/decision fusion (i.e., the use of majority voting) over the three best approaches.