Perceived personality
A carefully bounded construct in NIPG’s portfolio, and one where the safeguards matter more than usual because the everyday meaning of “personality” invites over-claiming.
What it is
Perceived personality state estimation in dyadic and small group interaction with deep learning methods (Fenech et al., arXiv) combines audio features, transcripts, and facial action units in a transformer-based model. It studies the personality expressed to, and perceived by, observers over short interaction segments, and explicitly distinguishes this from self-reported personality.
Safeguard. Perceived personality is not a direct measure of a person’s true personality, diagnosis, or current stress. It correlates with first impressions.
The ELEA result
In the ELEA collaborative task, group-average perceived-personality features predicted group performance more accurately than group-average self-reported Big Five features.
Safeguard. This result concerns collaborative group performance in one dataset. It is not diagnosis and not psychotherapy outcome.
Why it is nonetheless relevant here
It motivates two long-term research objectives, explicitly not expected clinical benefits:
- An adaptive avatar that responds to observed interaction patterns. → Avatar therapy
- Exploratory prediction of poor therapeutic fit or possible early discontinuation, so that clinicians can consider timely adaptation or an alternative therapy.
The probability of achieving either goal cannot yet be estimated. They may be stated as research objectives only.
Evaluation across external datasets
The work was evaluated across several external corpora — ChaLearn First Impressions, MULTISIMO, UDIVA, ELEA, AMI Meeting Corpus, MeMo — which is where the out-of-distribution and task-dependence evidence comes from. → Datasets
Related pages
Multimodal behaviour analysis · Avatar therapy · Datasets · NIPG · Editorial safeguards