Deep Learning for Sentiment Analysis (16 hours)

Teacher: Francesco Pugliese

In recent years, Sentiment Analysis (SA) has attracted significant attention in different areas of research and business
applications. Indeed, “sentiments” (emotions, motivations) can influence the opinions of many relevant actors, such as
product vendors, politicians, and the public opinion. Lately, Deep Learning (DL) models have been increasingly
employed in SA thanks to their automatic high-dimensional feature extraction capabilities. In this module, we will
understand and apply “Recurrent Convolutional Neural Networks” (RCNNs), that achieve “state-of-art” accuracy in
sentiment classification, to some concrete problems. We will also see applications of “Recursive Neural Tensor
Networks” (RNTNs) that reach an even better accuracy but require more pre-processing. We will also exploit an upside
of Deep Learning, the possibility of extracting keywords from the final max-pooling layer of models to create a corpus of domain-specific keywords for different applications: from keywords-driven sentiment analysis to an extractive topic-summarization. Hands-on use of the illustrated techniques, based on dedicated libraries for marchine learning such as Keras and Tensor Flow, applied to selected case studies.