A Hybrid Lexico-Transformer Model for Real-Time Emotion Detection in English Text

Abstract

Emotion detection in text is expressed as a crucial component of almost all artificial intelligence (AI) applications, so far it remains a challenging approach because of linguistic variety and real-time situations. This paper suggests DeepEmotion+, a hybrid approach which gathers a custom-built emotional lexicon with the transformer-based contextual learning in order to enhance both the accuracy and emotion classification speed. The proposed approach consists of two main pipeline stages, which include: Lexical-Preprocessing, where the text is tokenized, part-of-speech tagged, and enriched utilizing an extra domain-specific impact lexicon; and Transformer-Classification, where contextual embeddings with the lightweight transformer and lexicon-derived features are obtained through a novel Dynamic Fusion Module (DFM). The proposed approach validates its method on many datasets, illustrating an overall F1-score enhancement of about 3-5% compared with state-of-the-art studies in streaming situations and conditions. DeepEmotion+ consistently achieves an average accuracy of about 87%. In addition, the proposed approach ensures inference latencies below 50 ms per sentence on a CPU, enabling real-time deployment. These results express the underscored effectiveness and efficiency of DeepEmotion+ in practical text analysis.