EmoLASP Combines LLMs and Logic for Emotion Recognition.

This research introduces EmoLASP, a novel approach to emotion recognition in conversational AI. It addresses the limitations of standalone language models, such as instability and high costs associated with fine-tuning or processing long dialogue histories. EmoLASP integrates a language model with Answer Set Programming (ASP), a form of declarative reasoning, to predict Valence-Arousal-Dominance (VAD) scores. Experiments on the IEMOCAP dataset demonstrated that EmoLASP consistently outperforms language models used in isolation, even when those models are given no prior dialogue context. The most significant performance gains were observed with prompt-only LLMs, which EmoLASP utilizes without requiring any fine-tuning. However, for pre-trained language models that have been fine-tuned and provided with dialogue history, the additional reasoning component offered less incremental benefit. The EmoLASP framework highlights the potential of combining neural networks with symbolic reasoning to achieve more consistent emotion predictions, while simultaneously reducing the resource demands typically associated with large language models. This hybrid approach offers a path towards more efficient and robust emotional AI systems.