Article Preview
TopIntroduction
Artificial intelligence (AI) has numerous applications in various fields, including cloud computing (Bisht & Vampugani, 2022; Ilyas et al., 2022), intelligent systems (Casillo et al., 2022; Deveci et al., 2023), digital transformation (Gupta et al., 2023; Li et al., 2023), text detection (Yen et al., 2021; Zhang et al., 2023), and more. However, AI generally lacks the ability to express human emotions. Emotional intelligence is an important branch of artificial intelligence, which has been widely studied and explored in the field of natural language processing (NLP) (Barbosa et al., 2022; Chopra et al., 2022; Ismail et al., 2022). Emotional text generation, in particular, holds great potential for a variety of applications. Research shows that systems that can express emotions significantly improve user satisfaction (Prendinger & Ishizuka, 2005; Abo-Hammour et al., 2013; Arqub & Abo-Hammour, 2014). In the field of dialogue systems, some studies have improved the generated responses by endowing the dialogue system with increased empathy towards human users (Colombo et al., 2019; Abo-Hammour et al., 2014; Abu Arqub et al., 2012). The controlled emotional text generation model not only enables a more meaningful dialogue between AI agents and humans, but also aims to establish emotional connections with readers. This model proves beneficial for conversation therapy robots, as it can produce suitable emotional responses according to the user's psychological state (Sarivougioukas & Vagelatos, 2022).
Despite the diverse range of applications for emotional text generation, how to better integrate emotions into the generation model is still a difficult problem. Conventional methods for emotional text generation primarily rely on discourse templates and manual rules, frequently demonstrating limitations in addressing complex situations. Deep learning based methods have gained widespread usage in NLP tasks since the proposal of the recurrent neural network (RNN) Encoder-Decoder model by Cho et al. (2014). Vaswani et al. (2017) proposed the transformer architecture, which has significantly advanced NLP tasks by enabling large-scale language model training on massive datasets (Radford et al., 2019). Although deep learning-based methods have made significant progress in classification tasks (Chen et al., 2022), generating emotional text remains a challenging task, in contrast to the success achieved in sentiment classification (Yang et al., 2019). In particular, generation models based on transformer architecture can leverage a significant amount of unlabeled data for training. Nevertheless, this approach makes it challenging to change the attributes of the generated text without either modifying the model's architecture or utilizing specific attribute data for fine-tuning (Keskar et al., 2019; Zhang et al., 2018). Therefore, it becomes more difficult to incorporate emotional attributes into the pretrained model. However, most of the emotional text generation models today are still based on the Seq2Seq model of the RNN architecture, failing to take advantage of the latest pretrained transformer models such as GPT.