Current sentiment analysis techniques can recognize the emotions conveyed by a video. However they are hard to identify the elicited emotions of viewers. On the other hand, some atypical emotions, including refreshing, sympathy, choking and bloody violence, often are more applicable and instructive than typical emotions, like happy, sad, and angry. Bullet comment are the direct real-time responses from viewers to video content, making them reliable indicators of elicited emotions. We present AeerBERT, an atypical elicited emotion recognition model based on BERT, that simultaneously recognizes eight atypical emotions in bullet comment and identifies corresponding video clips. Our model leverages BERT’s intermediate hidden layer encodings to address the semantic ambiguity inherent in atypical emotions. Experimental results demonstrate that AeerBERT outperforms existing sentiment analysis models on accuracy, precision, recall, and F1 score.