Work

Deep multimodal transformers trained on vlogs reveal emergent group personalities during collaborative interactions

Year: 2022
Type: dissertation
Abstract: Group selection has emerged during evolution. In this case, natural selection acts at the level of the group. It manifests itself during group interactions, a regular part of daily life. Almost every aspect of a person’s life is affected by collaborating with other people. As it is such an important dynamic, so too is the ability to quantify it. Although groups in settings such as the workplace may be quantified by their task success, other kinds of groups do not have such a metric. Traditionally psychologists have defined the Big Five Personality Traits by which to classify individual behavior. The hypothesis is that during interaction such traits may appear at the group level and "group personality" may emerge. This work will introduce a method for quantifying groups in two-dimensional space. Features extracted from source separated speech, raw video, and transcripts will be input to a linear multimodal transformer to output a per-person personality estimation. The 5-dimensional personality points have high correlations and can be projected into two combinations, to the dimensions of "stability" and "plasticity". I used the ELEA database, a putative ’winter survival task’ for my studies. Results show unique group-level personality emerges in most of the cases.
Language: English
Open Access status: green
OpenAlex ID: W4282016558
Cites: 40
Cited by:
Related to: 10