Abstract
Studies of emotional facial expressions generally reveal consensus among human participants about the social meanings of only six to fifteen basic expressions. Here we argue that perceivers use sequences of these expressions as the basis for generating a much larger, richer vocabulary of emotion states. Participants reviewed combinations of eight most consensual, static basic facial expressions, presented as a sequence of two images out of 8x8=64 possible sequences. They were required to describe in one word the "state of mind" of the person whose sequence of images was presented. To explore the perceivers' vocabulary of emotion states, we relied on computational methods, utilizing word embedding methods (Global Vectors for Word Representation) adapted from the field of Natural Language Processing. Our findings reveal that the perceived meanings of the sequences of facial expressions were a weighted average of the single expressions comprising them, resulting in 22 new emotion states, in addition to the eight basic expressions used in the experiment. An interaction between the first and the second expression in each sequence, indicated that each facial expression influenced the perception of the other expression, as well as the interpretation of the sequence as a whole. We also found that the product (i.e., interaction) of the vectors representing two sequential facial expressions predicted the consensus among participants about whether the sequence is commonly seen or not in daily life. This result supports the notion that algebraic vector operations can predict human perception, shaped by past natural experience. Together, our findings suggest that the vocabulary of emotion states conveyed by facial expressions is rich and not limited, as it is the outcome of different combinations of these expressions that together creates a continuous space in which every emotion state is a weighted combination of the 8 basic emotions.
Meeting abstract presented at VSS 2018