Fig. 3: Model comparison and sensitivity.
From: Modelling individual and cross-cultural variation in the mapping of emotions to speech prosody
a, Model comparison using the WAIC. The models are arranged by their WAIC score, where lower WAIC values indicate a better model fit. The following models are shown, from right to left: the null model containing only intercepts; the base model estimating the global mapping; the in-group model estimating the interaction between country and language (we call this interaction ‘culture’); the corpus model from Fig. 2b; and the big model, which is the in-group model additionally modelling speaker and sex differences. The error bars are standard errors of the WAIC. b, Zoomed-in version of the black box in a, showing the WAIC of the in-group models modelling the group-level effect of countries, languages or the interaction of both. The icons are introduced in detail in Supplementary Methods 2. c, UPGMA-generated language tree from Beaufils and Tomin29. d, UPGMA-generated culture tree from Euclidian distances among the Hofstede30 dimensions. e, Confusion matrices predicting the dataset for the base, in-group and big models. Overall performance is expressed in UAR. Each cell contains a recall value. The recall values for each row are normalized and sum to 1. SUR, surprise; SAD, sadness; HAP, happiness; FER, fear; DIS, disgust; ANG, anger; NEU, neutral. All models in a,b can be explored using an interactive visualization; see http://mapping-emotions.pol.works.
