InstructERC: Reforming Emotion Recognition in Conversation with Multi-task Retrieval-Augmented Large Language Models

Abstract: The field of emotion recognition of conversation (ERC) has been focusing on separating sentence feature encoding and context modeling, lacking exploration in generative paradigms based on unified designs. In this study, we propose a novel approach, InstructERC, to reformulate the ERC task from a discriminative framework to a generative framework based on LLMs. InstructERC makes three significant contributions: (1) it introduces a simple yet effective retrieval template module, which helps the model explicitly integrate multi-granularity dialogue supervision information. (2) We introduce two additional emotion alignment tasks, namely speaker identification and emotion prediction tasks, to implicitly model the dialogue role relationships and future emotional tendencies in conversations. (3) Pioneeringly, we unify emotion labels across benchmarks through the feeling wheel to fit real application scenarios. InstructERC still perform impressively on this unified dataset. Our LLM-based plugin framework significantly outperforms all previous models and achieves comprehensive SOTA on three commonly used ERC datasets. Extensive analysis of parameter-efficient and data-scaling experiments provides empirical guidance for applying it in practical scenarios.

• The paper presents a novel generative framework using retrieval templates to integrate multi-granularity dialogue features for emotion recognition.
• It incorporates auxiliary tasks like speaker identification and emotion prediction to model interpersonal dynamics and forecast emotional trajectories.
• Experimental results on IEMOCAP, MELD, and EmoryNLP datasets show superior performance and efficient parameter scaling over traditional methods.

InstructERC: Reformulating Emotion Recognition in Conversations with LLMs

The field of emotion recognition in conversation (ERC) is pivotal for advancing human-computer interaction by enabling machines to understand nuanced emotional expressions in dialogue. Traditional approaches in ERC have mainly relied on separating sentence feature encoding from context modeling, often resulting in a lack of integrated designs capable of addressing both elements harmoniously. The study presented herein introduces InstructERC, a framework aimed at transforming the ERC task from a traditional discriminative methodology into a generative approach by leveraging the capabilities of LLMs.

Key Contributions and Methodological Advances

1. Generative Framework with Retrieval Templates: InstructERC introduces a novel retrieval template module designed to integrate multi-granularity dialogue supervision, which enriches the ERC task by unifying the output and input through a sequence-to-sequence (Seq2Seq) paradigm, a significant departure from previous discriminative frameworks.
2. Emotion Alignment Tasks: The researchers introduce two auxiliary tasks—speaker identification and emotion prediction tasks—which serve to enrich the modeling of speaker roles and project future emotional states in dialogues. These tasks aim to provide implicit cues about interpersonal dynamics and emotional trajectories within conversations.
3. Unified Emotion Labeling: InstructERC proposes a unified approach to emotion labeling across different benchmarks using the concept of the "feeling wheel." This alignment aims to increase the applicability and consistency of the model across various datasets.

Experimental Results

The paper reports that InstructERC significantly outperforms previous state-of-the-art models on three well-known ERC datasets: IEMOCAP, MELD, and EmoryNLP. This achievement is attributed to the model's ability to leverage LLMs’ generative frameworks, which provide a more robust foundation for language understanding and emotional nuance. Experimental evaluations also highlight the model's efficiency in parameter utilization and data scaling, suggesting it as a viable option for practical implementations.

Implications and Future Directions

The practical implications of InstructERC are substantial, particularly in the development of AI systems capable of empathetic interactions, which are increasingly important in domains such as virtual assistants, mental health diagnostics, and customer service. Theoretically, the shift from discriminative to generative frameworks in ERC tasks represents a significant step forward in how emotional intelligence is conceptualized within AI systems.

Future work could explore the integration of multi-modal data sources, such as visual and auditory signals, to further enhance the model's contextual awareness and emotional accuracy. Additionally, expanding InstructERC to support real-time and cross-cultural emotional interpretations will likely improve its applicability in global settings.

In conclusion, InstructERC not only reforms the methodologies typically applied in ERC tasks but also sets a new standard for how emotional recognition systems can be developed using advanced LLMs. The research provides a strong foundation for future advancements in the emotional intelligence of AI systems.