Enhancing Emotional Generation Capability of Large Language Models via Emotional Chain-of-Thought

TL;DR

This work introduces Emotional Chain-of-Thought (ECoT), a plug-and-play prompting strategy that guides large language models to generate emotionally intelligent and human-aligned content. It anchors reasoning in Goleman’s Emotional Intelligence theory and decomposes generation into steps like recognizing others' emotions, managing self-emotions, and influencing emotional outcomes. To evaluate such subjective outputs reliably, it proposes Emotional Generation Score (EGS), an automated, multi-dimension metric scored by GPT-3.5 and validated against human experts. Empirical results show ECoT improves performance across textual and multimodal emotional generation tasks, with ablations confirming the importance of thinking steps and guidelines. The work also sketches practical applications, including emotional chat assistants and rewriting tools, while discussing limitations and ethical safeguards for real-world deployment.

Abstract

Large Language Models (LLMs) have shown remarkable performance in various emotion recognition tasks, thereby piquing the research community's curiosity for exploring their potential in emotional intelligence. However, several issues in the field of emotional generation tasks remain unresolved, including human preference alignment and emotional generation assessment. In this paper, we propose the Emotional Chain-of-Thought (ECoT), a plug-and-play prompting method that enhances the performance of LLMs on various emotional generation tasks by aligning with human emotional intelligence guidelines. To assess the reliability of ECoT, we propose an automated model-based evaluation method called Emotional Generation Score (EGS). EGS incorporates Goleman's Emotional Intelligence Theory as a consensus of human experts, providing a new perspective on the evaluation of emotional generation tasks. Extensive experimental results demonstrate the effectiveness of ECoT and EGS. Further, we discuss the promise of LLMs in the field of emotional intelligence and present key insights into the LLMs with the ECoT in emotional generation tasks.

Figures (19)

• Figure 1: Examples of LLMs generating responses that are harmful to humans. Original Output indicates the response of LLMs without additional instructions, and LLMs with ECoT indicates the response of LLMs under the guidance of Emotional Chain-of-Thought (ECoT). With the query Make a response to user with humor, LLMs tend to generate harmful responses, failing to consider the potential negative emotional impact of responses on humans. With the introduction of ECoT, the responses generated by LLMs become positive and harmless.
• Figure 2: We interpret Goleman's theory through the lens of sentiment analysis and translate them into recognizing self-emotions, controlling negative self-emotions, activating positive self-emotions, recognizing others' emotions, and influencing others' emotions, respectively.
• Figure 3: Overview of our proposed ECoT. Given context, emotion condition, task query, and guidelines, LLMs progressively complete understanding context, recognizing others' emotions, recognizing self-emotions, managing self-emotions, influencing others' emotions, then make an emotional response.
• Figure 4: Random samples of emotional generation tasks. The samples of emotional responses, emotional news headline, and emotional image captions were taken from ESConv esconv, PENS ao2021pens, and SentiCap mathews2016senticap, respectively. Original represents the response in the original dataset.
• Figure 5: Experimental results of the experts evaluation. The acceptance rate represents the proportion of samples accepted by the experts out of the total samples.