Temporal Shift for Personality Recognition with Pre-Trained Representations

Personality, as a reflection of our inner psychological traits, profoundly influences our behavioral patterns, decision-making processes, and the way we interact with the world. Therefore, personality recognition has gradually become an important research direction in the fields of psychology and human-computer interaction. However, most previous works have focused on using pre-trained models to construct more complex models for feature extraction, neglecting the integration of multi-channel information and the fusion between channel information. To address the issue of multi-channel information integration, we propose and validate an innovative method based on multi-scale convolutional feature extraction and a temporal shift module for personality recognition in the speech modality. We use the wav2vec pre-trained model to extract features and enhance information fusion between channels by introducing multi-scale convolutional kernels and a temporal shift module. It expands the model's perception range and sensitivity to personality. Experimental results show that this method performs well in terms of accuracy (ACC) and concordance correlation coefficient (CCC).