Abstract
Current Emotion Recognition in Conversation (ERC) research follows a closed-domain assumption. However, there is no clear consensus on emotion classification in psychology, which presents a challenge for models when it comes to recognizing previously unseen emotions in real-world applications. To bridge this gap, we introduce the Unseen Emotion Recognition in Conversation (UERC) task for the first time and propose ProEmoTrans, a solid prototype-based emotion transfer framework. This prototype-based approach shows promise but still faces key challenges: First, implicit expressions complicate emotion definition, which we address by proposing an LLM-enhanced description approach. Second, utterance encoding in long conversations is difficult, which we tackle with a proposed parameter-free mechanism for efficient encoding and overfitting prevention. Finally, the Markovian flow nature of emotions is hard to transfer, which we address with an improved Attention Viterbi Decoding (AVD) method to transfer seen emotion transitions to unseen emotions. Extensive experiments on three datasets show that our method serves as a strong baseline for preliminary exploration in this new area.
Introduction
Emotion Recognition in Conversation (ERC) is a critical component in the development of empathetic artificial intelligence and sophisticated sentiment analysis tools. Despite its importance, current ERC models largely operate under a "closed-domain" assumption, meaning they are trained to recognize a fixed, pre-defined set of emotions such as joy, sadness, or anger.
In real-world scenarios, this assumption fails. Psychological research offers no single consensus on emotion classification, with theories ranging from six basic categories to over thirty complex states. When an AI encounters an emotion outside its training set—such as "peaceful" or "powerful"—it typically lacks the framework to categorize it correctly. This paper introduces the Unseen Emotion Recognition in Conversation (UERC) task, which requires models to identify novel emotion labels by transferring knowledge from known categories. To address this, the authors propose ProEmoTrans, a prototype-based framework designed for high generalization in zero-shot environments.
The ProEmoTrans Framework
ProEmoTrans addresses three primary hurdles in UERC: the difficulty of defining implicit emotions, the risk of overfitting during utterance encoding, and the challenge of transferring emotional transition dynamics.
LLM-enhanced Emotion Description (LED)
Traditional zero-shot models often rely on dry dictionary definitions to create emotion "prototypes." However, emotions in conversation are frequently expressed implicitly. To create richer, more representative prototypes, ProEmoTrans utilizes Large Language Models (LLMs) to generate sentences that describe the experience or manifestation of an emotion. For instance, instead of a literal definition for "Joy," the LED module might generate: "A bright smile spread across her face as she heard the news." These contextual examples allow the model to build more robust vector representations for matching against conversation utterances.
Gaussian Self-Attention (GAS)
Processing long conversations typically requires complex neural architectures like Graph Neural Networks (GNNs) or Recurrent Neural Networks (RNNs). While effective in closed-domain settings, these models often overfit to the specific emotional patterns of the training data, hindering their ability to generalize to new labels. ProEmoTrans introduces Gaussian Self-Attention (GAS), a parameter-free mechanism. GAS applies a Gaussian distribution to weight utterances based on the assumption that nearby context is more relevant than distant utterances. By removing learnable parameters in this stage, the model maintains the flexibility needed for unseen emotion recognition.
Attention Viterbi Decoding (AVD)
Conversations possess a Markovian property where the current emotional state is influenced by the previous one (e.g., a state of "Disgust" often transitions into "Anger"). ProEmoTrans employs an improved Attention Viterbi Decoding (AVD) method to leverage these transition patterns. By calculating the semantic similarity between seen and unseen emotion prototypes, the model can map learned transition flows onto entirely new emotion categories, improving the logical consistency of its predictions.
Experimental Setup and Results
The researchers evaluated ProEmoTrans using a cross-dataset approach across three major ERC datasets: IEMOCAP (scripted actor interactions), EmoryNLP (transcripts from Friends), and MELD (multi-party dialogue). By training on one dataset and testing on another with non-overlapping labels, the authors simulated a true "unseen" environment.
Performance Comparison
ProEmoTrans consistently outperformed both specialized ERC baselines and state-of-the-art LLMs in zero-shot scenarios:
- Superiority over LLMs: The model outperformed GPT-4o and DeepSeek-V3 across all experimental settings. In a transfer from EmoryNLP to IEMOCAP, ProEmoTrans achieved a weighted F1 score of 37.27%, significantly higher than DeepSeek-V3’s 25.69%.
- Ablation Analysis: The importance of the LED module was particularly pronounced; removing LLM-generated descriptions resulted in a 14.3% drop in average F1 score, confirming that high-quality, implicit prototypes are essential for zero-shot success.
- Efficiency: Despite its sophisticated components, the model remains computationally efficient with an inference time of approximately 6-9ms per sample, making it viable for real-time applications.
Conclusion and Future Work
ProEmoTrans establishes a strong baseline for the UERC task by demonstrating that "less is more" regarding parameterization and that LLMs are most effective when used to provide contextual examples rather than simple definitions.
However, limitations remain. The current framework relies on manually designed prompts for the LED module and is restricted to text-only analysis. Future research will explore automated prompt tuning and the integration of multi-modal cues—such as facial expressions and vocal tonality—to better capture the nuances of human emotion in complex conversations.
Summary updated