LLMs Generate Interpretable Explanations for Fine-Grained Social Media Sentiment Analysis

A recent research paper presents a method for analyzing sentiment about specific aspects in multimodal content—such as posts on social media that combine both images and text—while simultaneously generating human-readable explanations for its conclusions.

The Problem: Black-Box Sentiment Analysis#

Traditional approaches to multimodal aspect-based sentiment analysis (MABSA) rely on classification models that identify opinions at a fine-grained level but offer little transparency into their decision-making. Existing systems fuse textual and visual information to predict sentiment but typically cannot explain why they reached their conclusions, limiting their usefulness for content understanding and trust.

A Generative, Explainable Alternative#

Rather than treating MABSA as a classification task, the researchers reformulated it as a generative problem. Their framework uses multimodal large language models (MLLMs) to simultaneously predict aspect-level sentiment and produce natural language explanations, leveraging a prompt-based approach that treats both as interconnected outputs.

Dependency Syntax Guidance#

To strengthen the model's ability to reason about specific aspects, the team introduced a dependency-syntax-guided sentiment cue strategy. This technique:

• Prunes and converts aspect-centered dependency syntax trees into textual form
• Helps the model distinguish sentiment across different aspects
• Enhances the model's capacity to generate aspect-grounded explanations

The researchers used MLLMs to construct explanation-augmented datasets for fine-tuning, enabling the system to learn to produce coherent justifications alongside sentiment predictions.

Results#

Experiments demonstrated that the approach not only improved overall sentiment classification accuracy but also generated explanations that were both logically sound and tethered to specific aspects discussed in the multimodal content.

Source: arXiv:2601.06848