Trustworthy Multimodal Aspect-Based Sentiment Analysis with Vision-Language Models: Aspect-Guided Reasoning, Evidence Grounding, and Explainable Sentiment Understanding

Abstract: Multimodal Aspect-Based Sentiment Analysis (MABSA) has emerged as a critical research area that aims to identify aspect-specific opinions from multimodal data such as text and images. Traditional sentiment analysis methods often fail to capture fine-grained opinions associated with individual aspects and struggle to leverage complementary visual information. Recent advances in Vision-Language Models (VLMs) and Multimodal Large Language Models (MLLMs), including CLIP, BLIP, LLaVA, and GPT-4V-inspired frameworks, have significantly enhanced multimodal understanding through cross-modal reasoning and knowledge transfer. Despite remarkable performance improvements, challenges related to trustworthiness, explainability, visual grounding, hallucination, fairness, and robustness remain unresolved. This review provides a comprehensive analysis of recent developments in trustworthy MABSA, emphasizing aspect-guided reasoning, evidence grounding, explainable sentiment understanding, and emerging Vision-Language architectures. Furthermore, we discuss benchmark datasets, evaluation methodologies, open challenges, and future research opportunities. The review highlights how grounded multimodal reasoning can establish reliable and transparent sentiment analysis systems suitable for real-world applications in e-commerce, healthcare, tourism, social media monitoring, and decision intelligence.