In this paper, we present a method of predicting emotions from multi-label conversation transcripts. The transcripts are from a movie dialog corpus and annotated partly by 3 annotators. The method includes building an emotion lexicon bootstrapped from Wordnet following the notion of Plutchik's basic emotions and dyads. The lexicon is then adapted to the training data by using a simple Neural Network to fine-tune the weights toward each basic emotion. We then use the adapted lexicon to extract the features and use them for another Deep Network which does the detection of emotions in conversation transcripts. The experiments were conducted to confirm the effectiveness of the method, which turned out to be nearly as good as a human annotator.
展开▼
