Deep Learning Approach to predict Emotional Effect in Conversations

The recent explosion of human-chatbot interactions, encompassing everything from online insurance purchases to grocery deliveries, has created a daily reality of pervasive two-party conversations mediated by artificial intelligence. With this exponential growth, the need for emotionally aware systems becomes increasingly paramount. While existing Emotion Recognition from Conversations (ERC) approaches excel at identifying emotions within individual utterances, they fall short in capturing the holistic emotional journey experienced by participants throughout the entire conversation. To address this limitation, we present our work on 'Emotional Effect,' which delves deeper by analyzing the post-conversational emotional state of each participant. This approach provides a more nuanced understanding of the emotional impact of human interactions, paving the way for the development of emotionally intelligent systems that foster positive and engaging user experiences.A fundamental obstacle in research of emotional effect has been the scarcity of high-quality, tagged data. This paper proposes a novel method for data oversampling utilizing GPT-3.5, demonstrably the most effective approach for conversational data augmentation. Leveraging the expressive power of BERT embeddings and the sentiment analysis capabilities of NRCLex, we explore the prediction of "joy" as an emotional effect in conversations. This paper presents the outcomes of our investigation, assessing the model's F1-Score (60%) and offering interpretations of its predictions.