QLoRA Fine-Tunes LLMs for Multimodal, Multilingual Opinion Extraction

Large language models (LLMs) have transformed semantic analysis, but extracting concise core opinions for Science and Technology Intelligence (STI) from vast, noisy information streams remains challenging. Off-the-shelf models often struggle with filtering irrelevant information and producing reliable structured outputs, especially in zero-shot multilingual and multimodal contexts. To address these issues, this research introduces a multimodal core-opinion extraction framework. It leverages visual evidence as a contextual anchor to enhance textual judgment. The approach involves fine-tuning VideoLLaMA2 (VL2) and VideoLLaMA2.1 (VL2.1) base models using Quantized Low-Rank Adaptation (QLoRA) on a specially curated dataset of 2,194 multilingual and multimodal samples. Under an Image-Augmented setting, the fine-tuned VL2.1 model generates structured JSON core-opinion outputs with notable improvements: 64.98% Precision, 42.15% Recall, 51.14% F1-score, and 74.00% sample-level accuracy. This represents a substantial gain over the zero-shot VL2.1 setting, particularly for languages like Spanish and Russian, where F1-scores rose from 4.83% and 0.45% to 46.05% and 51.93% respectively. The framework also includes a Fuzzy Cumulative Prospect Theory-based triage module for assessing case-level value, providing a crucial signal for downstream STI screening.