ECGText: Human-Centric Text Generation with Enhanced Emotional Intelligence
Biswajit Nath Roy, Soumik Maity, Kabir Raj Singh, Pankaj Chowdhury, Arijit Das, Diganta Saha · 2024
The recent improvements in the natural language processing (NLP) field have made it possible to create extremely powerful models such as Generative Pre-trained Transformers (GPT), used for generating human-sounding, contextually relevant text. Nevertheless, these types of models often fail to convey the proper emotional fidelity that we need for human communication. We propose ECGText (Emotional Controlled Generation of Text), a new framework that implements GPT to respond Emotion-controlled prompt with the assistance of RoBERTa-based GoEmotions model for emotion recognition in order to mitigate this limitation. ECGText generates emotionally aligned text while maintaining coherence and the author’s writing style. The model supports multiple languages and is designed for applications requiring emotionally intelligent text. We evaluated ECGText using various datasets, including Obama speeches, film reviews, and Rabindranath Tagore’s works. The results demonstrate that ECGText achieves superior performance in text diversity, coherence, and emotional alignment compared to baseline models. This work represents a significant step toward developing emotionally intelligent AI systems, with potential applications in affective computing and human-AI interaction.