A Multi-channel BiLSTM-CNN Model for Multilabel Emotion Classification of Informal Text
Zahra Rajabi, Amarda Shehu, Özlem Uzuner
Abstract
State-of-the-art research on emotion classification from text primarily focuses on binary or ternary classification. Yet, humans express a variety of emotions. Here we approach the classification of emotions from short, informal text as a multi-class problem, employing popular psychology models of basic and advanced human emotions. We account for imbalanced datasets differing in annotation schema, psychological models considered, and number of annotated emotions. We show that a multi-channel, multi-filter CNN-BiLSTM outperforms existing models, achieving 82.3% accuracy on the multi-label SemEval18-EC dataset.
Topics & Concepts
Computer scienceEmotion classificationSchema (genetic algorithms)Artificial intelligenceVariety (cybernetics)Classifier (UML)AnnotationBinary classificationFilter (signal processing)Emotion recognitionMachine learningNatural language processingSupport vector machineComputer visionSentiment Analysis and Opinion MiningAdvanced Text Analysis TechniquesText and Document Classification Technologies