Fine-Tuning Multimodal Models for Multilingual Event and Opinion Extraction in Science and Technology Intelligence

Multimodal large language models (MLLMs) have demonstrated high performance in information extraction, but their ability to conduct multilingual, multimodal event extraction (EE) and opinion extraction (OE) for science and technology intelligence (STI) is yet to be proven. This study examines VideoLLaMA2 (VL2) and VideoLLaMA2.1 (VL2.1) on a manually annotated dataset of 5, 728 E E and 2,194 OE samples for STI across English, Chinese, Spanish, and Russian, using text, image, and video inputs. Zero-shot results show VL2 achieving 46.11% for OE, 28.40% for EE trigger (Tr), and 23.89% for EE argument (Arg), and VL2.1 achieving 47.39% for OE, 24.76% for EE trigger (Tr), and 20.88% for EE argument (Arg), improved by prompt techniques: Chain-of-Thought, Tree-of-Thought and 4 shots prompting. While fine-tuning delivers the highest gains with VL2.1 achieving 74.43% accuracy in opinion extraction (27.04% baseline improvement) and 65.48% for event trigger identification (40.72% baseline improvement). Fine-tuning outperformed prompt engineering, offering a robust framework for multimodal, multilingual STI extraction.