The "27 emotions" classification is the kind of thing that people come up when they want to do ML experiments forever and never face failure.<p>If the 27 emotions were uniformly distributed, you could do better than chance by guessing the right emotion 1 time in 20 (5% of the time!) but that would be completely useless from an engineering viewpoint.<p>If "Happy"/"Sad" were not good enough for you, try one of these<p><a href="https://en.wikipedia.org/wiki/Emotion_classification#Dimensional_models_of_emotion" rel="nofollow">https://en.wikipedia.org/wiki/Emotion_classification#Dimensi...</a>