Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval (2410.23041v1)

Abstract: As LLMs exhibit a high degree of human-like capability, increasing attention has been paid to role-playing research areas in which responses generated by LLMs are expected to mimic human replies. This has promoted the exploration of role-playing agents in various applications, such as chatbots that can engage in natural conversations with users and virtual assistants that can provide personalized support and guidance. The crucial factor in the role-playing task is the effective utilization of character memory, which stores characters' profiles, experiences, and historical dialogues. Retrieval Augmented Generation (RAG) technology is used to access the related memory to enhance the response generation of role-playing agents. Most existing studies retrieve related information based on the semantic similarity of memory to maintain characters' personalized traits, and few attempts have been made to incorporate the emotional factor in the retrieval argument generation (RAG) of LLMs. Inspired by the Mood-Dependent Memory theory, which indicates that people recall an event better if they somehow reinstate during recall the original emotion they experienced during learning, we propose a novel emotion-aware memory retrieval framework, termed Emotional RAG, which recalls the related memory with consideration of emotional state in role-playing agents. Specifically, we design two kinds of retrieval strategies, i.e., combination strategy and sequential strategy, to incorporate both memory semantic and emotional states during the retrieval process. Extensive experiments on three representative role-playing datasets demonstrate that our Emotional RAG framework outperforms the method without considering the emotional factor in maintaining the personalities of role-playing agents. This provides evidence to further reinforce the Mood-Dependent Memory theory in psychology.

• The paper presents the Emotional RAG framework leveraging emotional memory retrieval based on mood-dependent memory to enrich agent interactions.
• The methodology encodes both semantic and emotional states using advanced vector models and employs dual retrieval strategies for improved query relevance.
• Experimental results show enhanced personalization and consistent character portrayal, notably in systems like ChatGLM-6B and Qwen-72B, outperforming traditional methods.

Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval

Introduction

The paper "Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval" introduces an advanced framework designed to augment LLMs with emotional memory retrieval capabilities. The Emotional RAG framework is motivated by the need for LLMs to not only retrieve semantically relevant memories but also memories consistent with emotional states, thus providing role-playing agents that can emulate more nuanced, human-like responses. This approach leverages psychological theories, specifically Mood-Dependent Memory, to enhance interactive conversations with emotional depth.

Figure 1: The overview architecture of Emotional RAG framework. It contains four components: the query encoding component, the memory encoding component, the emotional retrieval component, and the response generation component.

Methodology

The proposed framework, Emotional RAG, incorporates several key components to retrieve and utilize emotional memory, enhancing an agent's conversational capabilities.

Overview Architecture

The Emotional RAG framework comprises four core components:

1. Query Encoding Component: Encodes both semantic and emotional state vectors of incoming queries.
2. Memory Encoding Component: Stores and encodes historical conversation data similarly in terms of semantic and emotional vectors.
3. Emotional Retrieval Component: Retrieves memory fragments based on both semantic relevance and emotional congruity, utilizing multiple retrieval strategies, including combination and sequential strategies.
4. Response Generation Component: Utilizes a prompt that integrates the query, character profiles, and retrieved emotional memory to generate responses, using templates like those depicted below.

   Figure 2: Emotion scoring prompt template in LLMs.

Technical Implementation

The framework employs advanced vector embedding models and utilizes the Semantic Enhanced Embedding (bge-base-zh-v1.5) and GPT-3.5 for emotion scoring. Emotional retrieval is implemented through two retrieval strategies—combination (C-A/C-M) and sequential (S-S/S-E)—to determine the emotional and semantic relevance of memory fragments, ensuring that the emotion-congruity criterion is preserved.

Figure 3: An example of response generation prompt template in the CharacterEval dataset.

Experimental Evaluation

The effectiveness of Emotional RAG was evaluated across three datasets: InCharacter, CharacterEval, and Character-LLM, using BFI and MBTI as evaluation metrics. The results indicated that:

• Performance Improvements: Emotional RAG consistently outperformed traditional RAG methods, particularly in full personality type evaluations, by leveraging emotional retrieval to maintain and reflect character personalities more accurately.
• Robustness Across Models: While improvements were seen in all models, the largest gains were in systems like ChatGLM-6B and Qwen-72B, suggesting the importance of emotional congruity when semantic capabilities are otherwise robust.

  Figure 4: An example of the prompt template for dimension Extroversion (E) vs. Introversion (I) in MBTI evaluation.

Analysis and Discussion

RAG Strategy Analysis

The paper analyzes the impact of different retrieval strategies in Emotional RAG, indicating that varying approaches have advantages depending on the evaluation criteria used (e.g., BFI vs. MBTI). The preference for sequential emotional prioritization in particular contexts highlights the role of emotion in personalization tasks.

Case Studies

Case studies illustrate how Emotional RAG enables agents to produce responses better aligned emotionally with queries, resulting in interactions that are more coherent and contextually appropriate when managing emotionally loaded dialogues.

(Figure 5 and Figure 6)

Figure 5: Different RAG strategies on BFI evaluation.

Figure 6: Two examples that Emotional RAG generates better results.

Implications and Future Work

This research underscores the significant potential of integrating emotional awareness into memory retrieval processes for AI agents. The development of the Emotional RAG framework is a step toward achieving nuanced, empathetic AI-driven interactions across domain applications like virtual customer service and interactive entertainment. Future research could focus on optimizing memory architecture and exploring adaptive learning mechanisms to further refine the emotional alignment of role-playing agents.

Conclusion

Emotional RAG framework represents a significant advancement in role-playing agents, demonstrating that incorporating emotional retrieval strategies significantly enhances LLM's ability to generate personified, emotionally resonant conversations. This augmentation validates mood-dependent memory from psychology, paving the way for more sophisticated solutions in human-computer interaction domains.