Evaluating Multimodal Large Language Models on Spoken Sarcasm Detection
The study systematically tested Large Language Models (LLMs) and multimodal LLMs for detecting sarcasm in English and Chinese using various settings like zero-shot and fine-tuning. Findings show that audio-based models are best alone, but combining modalities like text-audio or audio-vision significantly boosts performance, especially when using parameter-efficient fine-tuning methods. This highlights the power of integrating different data types for better sarcasm understanding across languages.