Beyond Information: Is ChatGPT Empathetic Enough?

Proceedings of Recent Advances in Natural Language Processing, pages 159–169 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_018 159 Beyond Information: Is ChatGPT Empathetic Enough? Ahmed Belkhir Universit´e du Qu´ebec `a Montr´eal belkhir.ahmed@courrier.uqam.ca Fatiha Sadat Universit´e du Qu´ebec `a Montr´eal sadat.fatiha@uqam.ca Abstract This paper aims to explore and enhance Chat- GPT’s abilities to generate more human-like conversations by taking into account the emo- tional state of the user. To achieve this goal, a prompt-driven Emotional Intelligence is used through the empathetic dialogue dataset in or- der to propose a more empathetic conversa- tional language model. We propose two al- tered versions of ChatGPT as follows: (1) an emotion-infused version which takes the user’s emotion as input before generating responses using an emotion classifier based on ELECTRA (Clark et al., 2020); and (2) the emotion adapt- ing version that tries to accommodate for how the user feels without any external component. By analyzing responses of the two proposed al- tered versions and comparing them to the stan- dard version of ChatGPT, we find that using the external emotion classifier leads to more fre- quent and pronounced use of positive emotions compared to the standard version. On the other hand, using simple prompt engineering to take the user emotion into consideration, does the opposite. Finally, comparisons with state-of- the-art models highlight the potential of prompt engineering to enhance the emotional abilities of chatbots based on large language models. 1 Introduction Conversational agents have become increasingly popular in recent years, with applications ranging from customer service (Ando and Zhang, 2005) to mental health therapy (Abd-Alrazaq et al., 2019). However, while these agents have the potential to provide information in natural language, their current abilities to generate human-like and empa- thetic conversations are limited (Rapp et al., 2021; Belainine et al., 2020b,a). To address this challenge, this study explores the emotional abilities of ChatGPT in generating empa- thetic responses. Specifically, we investigate the ef- fectiveness of incorporating external emotion clas- sifiers using prompt engineering to take the user’s emotional state into account when generating re- sponses. This study is motivated by the fact that emotions play a crucial role in human communi- cation, and empathetic responses are essential for building rapport and trust in human-machine inter- actions (Chen et al., 2021). In customer service for instance, it was shown that up to 40% of con- sumers’ requests are rather emotional without spe- cific informational intents (Xu et al., 2017). Thus, we compare standard ChatGPT that generates re- sponses to simple conversation prompts from the Empathetic Dialogues dataset (Ma et al., 2020) to two slightly modified versions with prompt engi- neering. The first one is an emotion-infused version that takes the user emotion as an additional input before generating responses using an ELECTRA- based emotion classifier; while the second, emotion adapting version tries to consider how the user feels without any external component. Our study adds to the expanding literature on con- versational agents and emotional intelligence and its results have implications for the design and de- velopment …

Content truncated to 500 words