Using Natural Language Processing to Characterize the Phenomenology of Deja Vu

Abstract

Around two decades ago, cognitive psychologists began empirically studying déjà vu—the odd sensation that one is re-experiencing something from the past while being certain of the situation’s novelty. Prior work, investigating déjà vu’s subjective phenomenology by soliciting retrospective reports in a survey, uncovered déjà vu was relatively positive in its emotionality, and involved language pertaining to places, consistent with past survey research suggesting that scenes are the most common elicitor of déjà vu. However, there are well-documented limitations to relying on retrospective memories of an experience for studying it. The present thesis aimed to instead examine language that occurred around the moment that a déjà vu experience happened. Natural language processing (NLP) was performed on transcribed spoken language that occurred in a Think Aloud protocol during a Virtual Reality implementation of a well-established method for studying déjà vu—the Virtual Tour Method. Machine learning (ML) classification models were trained to distinguish déjà vu vs. non-déjà vu reports and found that déjà vu was characterized by increased filler words and recollective confabulation (i.e. recollection of incorrect episodic details). In line with prior work, trials where participants reported déjà vu were more positive, but that was not an important feature in the ML models predictions. Additionally, spatial features of the environment were not influential to the models’ predictions. These results help characterize more of the déjà vu experience and open a new direction for exploration into the relationship between internal attention and déjà vu.