Using natural language processing to characterize the phenomenology of déjà vu
Open MIND January 1, 2026 Sarah Nicole Horne, Anne M. Cleary, Joshua Prasad et al.
Déjà vu—the strange feeling of reliving a past experience despite knowing the situation is new—has been studied mainly through retrospective surveys, which have limitations. This thesis used natural language processing on spoken language from a Think Aloud protocol during a Virtual Reality version of the Virtual Tour Method, a standard technique for eliciting déjà vu. Machine learning models distinguished déjà vu from non-déjà vu reports, finding that déjà vu involved more filler words and recollective confabulation (recalling incorrect episodic details). Consistent with prior work, déjà vu reports were more positive emotionally, but this was not a key predictor in the models. Spatial features of the environment were also not influential. These results characterize more of the déjà vu experience and suggest a link between internal attention and déjà vu.