End-to-End Hierarchical Approach for Emotion Detection in Short Texts

Abstract

AbstractTextual emotion analysis aims to identify and recognize a set of predefined emotions from the text (e.g., sentences, documents). Vast amounts of textual data are created and distributed daily, for example, in news reports and social media, especially microblogging services. Understanding emotion in short text snippets has become an important area of study due to its proliferation in daily life. This paper proposes a novel hierarchical approach for detecting emotion from short texts (microblogs) and conducts a broad set of experiments. Based on the observation that sentiment classification can get up to 80–90% accuracy, we develop a hierarchical end-to-end emotion classifier in which the first layer is the polarity detector and the second layer is the emotion detector. The proposed hierarchical approach can incorporate state-of-the-art neural models, and we have experimented with a BiLTSM-based self-attention model, a BiGRU model, and a BERT-based model for hierarchical emotion classification. Our approaches are evaluated on four large datasets, demonstrating statistically significant improvements.KeywordsEmotion detectionDeep learning modelsLanguage understanding