Table 1.

Pearson correlations and confidence intervals for unimodal and multimodal models. The top panel shows the performance when v_a corresponds to the distributional linguistic model, while the middle panel v_a corresponds to the word association baseline. The bottom panel corresponds to purely experiential model where v_a corresponds to the affective model and is added for completeness. In each panel, the unimodal columns show the performance of that model (v_a) as well as the two experiential models (v_b) on either the concrete or the abstract words. The best‐fitting multimodal models combining v_a and v_b were found by optimizing the correlation for mixing parameter β and are shown in column $r_{v_{\mathit{ab}}}$. The improvement due to adding experiential information ($r_{v_{\mathit{ab}}}‐r_{v_a}$) is shown in column Δr

Dataset	$n$	va = Distributional linguistic model
		Unimodal					Multimodal
		$r_{v_a}$	CI95	vb	$r_{v_b}$	CI95	β	$r_{v_{\mathit{ab}}}$	CI95	Δr	CI95
Concrete	300	.64	[0.57, 0.70]	Visual	.67	[0.60, 0.73]	.48	.75	[0.70, 0.80]	.12	[0.07, 0.17]
Concrete	300	.64	[0.57, 0.70]	Affect	.21	[0.10, 0.32]	.50	.68	[0.62, 0.74]	.04	[0.02, 0.08]
Abstract	300	.62	[0.54, 0.68]	Affect	.51	[0.43, 0.59]	.58	.74	[0.69, 0.79]	.13	[0.08, 0.19]

Dataset	$n$	va = Word association model
		Unimodal					Multimodal
		$r_{v_a}$	CI95	vb	$r_{v_b}$	CI95	β	$r_{v_{\mathit{ab}}}$	CI95	Δr	CI95
Concrete	300	.76	[0.71, 0.80]	Visual	.67	[0.60, 0.73]	.35	.81	[0.77, 0.85]	.05	[0.03, 0.08]
Concrete	300	.76	[0.71, 0.80]	Affect	.21	[0.10, 0.32]	.38	.78	[0.74, 0.82]	.02	[0.00, 0.05]
Abstract	300	.82	[0.78, 0.86]	Affect	.51	[0.43, 0.59]	.05	.82	[0.78, 0.86]	.00	[−0.01, 0.01]

Dataset	$n$	va = Affective model
		Unimodal					Multimodal
		$r_{v_a}$	CI95	vb	$r_{v_b}$	CI95	β	$r_{v_{\mathit{ab}}}$	CI95	Δr	CI95
Concrete	300	.21	[0.10, 0.32]	Visual	.67	[0.60, 0.73]	.45	.73	[0.67, 0.77]	.52	[0.41, 0.63]

Note that the confidence intervals for Δr are based on testing significant differences for dependent overlapping correlations based on Zou (2007). This approach increases the power to detect an effect compared to Fisher's r to z procedure which assumes independence.