Emotion Identification in Italian Opera

This work aims to develop classification models able to automatically perform the task of emotion identification on Italian arias. These models enable the musicologists and the public interested in opera to investigate the emotion of Italian aria in a systematical way. An aria can be seen as an independent unit of opera that is sung by one character. Each aria contains 1 to 8 verses. Considering an aria may transmit more than one emotion, a lower level granularity is adopted: the identification of the emotion transmitted at the verse level. On the basis of a manually labelled corpus comprised of 2,500 aria verses with their corresponding emotion, the �first part of this work investigates different text representations and classification approaches. Building on the results of the exploration in the �first part, the second part investigates emotion identification at the aria level. The size of supervised data is expanded by means of self-learning. The verse-level annotation is converted into aria-level annotation and each aria is assigned up to two emotion labels. I experimented with pre-trained character trigram embeddings and convolutional neural network. For the emotion identification at the verse level, the combination of character trigram based TF-IDF and neural network with 2 hidden layers outperformed other combinations, achieving an accuracy of 0.47 on the test set. As for the emotion identification at the aria level, a convolutional neural network combined with character trigram based embeddings developed based on a corpus of Italian arias achieved an accuracy of 0:68.