Enriching Affect Analysis Through Emotion and Sarcasm Detection

Affect detection from text is the task of detecting affective states such as sentiment, mood and emotions from natural language text including news comments, product reviews, discussion posts, tweets and so on. Broadly speaking, affect detection includes the related tasks of sentiment analysis, emotion detection and sarcasm detection, amongst others. In this dissertation, we seek to enrich textual affect analysis from two perspectives: emotion and sarcasm. Emotion detection entails classifying the text into fine-grained categories of emotions such as happiness, sadness, surprise, and so on, whereas sarcasm detection seeks to identify the presence or absence of sarcasm in text. The task of emotion detection is particularly challenging due to limited number of resources and as it involves a greater number of categories of emotions in which to undertake classification, with no fixed number or types of emotions. Similarly, the recently proposed task of sarcasm detection is complicated due to the inherent sophisticated nature of sarcasm, where one typically says or writes the opposite of what they mean. This dissertation consists of five contributions. First, we address word-emotion association, a fundamental building block of most, if not all, emotion detection systems. Current approaches to emotion detection rely on a handful of manually annotated resources such as lexicons and datasets for deriving word-emotion association. Instead, we propose novel models for augmenting word-emotion association to support unsupervised learning which does not require labeled training data and can be extended to flexible taxonomies of emotions. Second, we study the problem of affective word representations, where affectively similar words are projected into neighboring regions of an n-dimensional embedding space. While existing techniques usually consider the lexical semantics and syntax of co-occurring words, thus rating emotionally dissimilar words occurring in similar contexts as highly similar, we integrate a rich spectrum of emotions into representation learning in order to cluster emotionally similar words closer, and emotionally dissimilar words farther from each other. The generated emotion-enriched word representations are found to be better at capturing relevant features useful for sentence-level emotion classification and emotion similarity tasks. Third, we investigate the problem of computational sarcasm detection. Generally, sarcasm detection is treated as a linguistic and lexical phenomena with limited emphasis on the emotional aspects of sarcasm. In order to address this gap, we propose novel models of enriching sarcasm detection by incorporating affective knowledge. In particular, document-level features obtained from affective word representations are utilized in designing classification systems. Through extensive evaluation on six datasets from three diverse domains of text, we demonstrate the potential of exploiting automatically induced features without the need for considerable manual feature engineering. Motivated by the importance of affective knowledge in detecting sarcasm, the fourth contribution of this thesis seeks to dig deeper and study the role of transitions and relationships between different emotions in order to discover which emotions serve as more informative and discriminative features for distinguishing sarcastic utterances in text. Lastly, we show the usefulness of our proposed affective models by applying them in a non-affective framework of predicting the helpfulness of online reviews.