Granular Emotion Detection for Multi-class Sentiment Analysis in Social Media
Sentiment analysis for text classification generally refers to assessing the polarity of the emotional context of written text, whether in a binary (e.g. positive or negative) or trinary (e.g. positive, neutral, or negative) state. Granular emotion detection is a more specialized form of sentiment analysis, wherein we move from predicting sentiment polarity to detecting specific classes of emotions within text (e.g. happy, sad, anger, love, hate, etc.), whether that context is a reflection of the author's own emotional state or the emotional state the author intended to convey. Granular emotion detection is broadly applicable to the business world, with common applications in customer satisfaction and retention, as well as studies of marketing effectiveness. Other applications include attempting to identify angry people based on their social media posts and prevent them from committing acts of violence. Current approaches to multi-class emotion classification show mixed or limited results, and improving accuracy for multiple classes of emotions is an open research challenge. Moreover, many modern application contexts align more directly with social media content or have a shorter format more akin to social media, where texts often bend or violate standard language conventions. Overall, understanding emotion detection in social media (EMDISM) contexts is an open challenge.To address the challenge of granular emotion detection in social media text, I have investigated ensemble approaches that combine a variety of individual classifiers to address tradeoffs in performance. This involved first investigating EMDISM performance for individual traditional machine learning (ML), deep learning (DL), and transformer learning (TL) classifiers. Based on this analysis, the second stage investigated the creation of ensembles of the most accurate classifiers across these general classes which offer comparatively improved performance. The approaches were evaluated based on a large Twitter dataset with more than 1.2M samples and encompassing seven discrete emotions. I provide results and analysis for each classifier I considered as well as the most accurate ensembles I created from the most accurate singleton classifiers. Results show that the proposed ensemble approaches - simple voting, weighted voting, cascading, and cascading/switching - improve upon the state of the art for average accuracy, weighted precision, weighted recall, and weighted f-measure as compared to the most accurate single classifier for EMDISM.