Associating specific emotions to short sequences of texts
We used BERT to classify short texts into one of the seven emotions, as opposed to the typical binary (positive/negative) or ternary (positive/negative/neutral) classes.
Our training and testing dataset are the International Survey on Emotion Antecedents and Reactions (ISEAR), consisting of 1096 records with labeled emotions of seven classes: anger, disgust, fear, joy, sadness, shame, and guilt.
We have achieved an overall accuracy of 69% with a macro F1, Recall, and Precision of 0.69. Moreover, this model indicates a good balance of sensitivity and specificity overall. Also, our results have been pretty close to that best metrics reported for the ISEAR dataset, which is 67%.
Contributors: Aiman Haider, Maobin Guo, Pranav Manjunath, Xinyi Pan
Institution: Duke University