| Description |
Over the last decade, social media has emerged as a revolutionary platform for informal communication and social interactions among people. Publicly expressing thoughts, opinions, and feelings is one of the key characteristics of social media. In this dissertation, I present research on automatically acquiring knowledge from social media that can be used to recognize people's affective state (i.e., what someone feels at a given time) in text. This research addresses two types of affective knowledge: 1) hashtag indicators of emotion consisting of emotion hashtags and emotion hashtag patterns, and 2) affective understanding of similes (a form of figurative comparison). My research introduces a bootstrapped learning algorithm for learning hashtag in- dicators of emotions from tweets with respect to five emotion categories: Affection, Anger/Rage, Fear/Anxiety, Joy, and Sadness/Disappointment. With a few seed emotion hashtags per emotion category, the bootstrapping algorithm iteratively learns new hashtags and more generalized hashtag patterns by analyzing emotion in tweets that contain these indicators. Emotion phrases are also harvested from the learned indicators to train additional classifiers that use the surrounding word context of the phrases as features. This is the first work to learn hashtag indicators of emotions. My research also presents a supervised classification method for classifying affective polarity of similes in Twitter. Using lexical, semantic, and sentiment properties of different simile components as features, supervised classifiers are trained to classify a simile into a positive or negative affective polarity class. The property of comparison is also fundamental to the affective understanding of similes. My research introduces a novel framework for inferring implicit properties that 1) uses syntactic constructions, statistical association, dictionary definitions and word embedding vector similarity to generate and rank candidate properties, 2) re-ranks the top properties using influence from multiple simile components, and 3) aggregates the ranks of each property from different methods to create a final ranked list of properties. The inferred properties are used to derive additional features for the supervised classifiers to further improve affective polarity recognition. Experimental results show substantial improvements in affective understanding of similes over the use of existing sentiment resources. |
