Teaching machines about emotions

Abstract

Artificial intelligence algorithms are becoming an increasingly important part of human life with many chat bots and digital personal assistants now interacting directly with us through natural language. Such human-computer interaction can be made more useful by enriching the underlying algorithms with a detailed sense of emotion. In my thesis I propose new ways to detect, encode and modify emotional content in text. First, I show how we can leverage the vast amount of texts on social media with emojis to train a classifier that can accurately detect various kinds of emotional content in text. Secondly, I introduce a state-of-the-art domain adaptation method that is explicitly designed to tackle issues occurring in the messy real-world text data that existing NLP methods struggle with. Lastly, I propose a new algorithm that could be used to decompose text inputs into disentangled representations and then manipulate these representations in a controlled manner to obtain a modified version of the input.