Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Explainable Multimodal Aspect-Based Sentiment Analysis with Dependency-guided Large Language Model

Zhongzheng Wang, Yuanhe Tian, Hongzhi Wang, Yan Song

TL;DR

The paper reframes multimodal aspect-based sentiment analysis as a generative task that jointly outputs sentiment labels and natural language explanations using multimodal large language models. It introduces a dependency-syntax guided cue strategy to prune aspect-centered dependencies and textualizes them as prompts, enhancing aspect-aware reasoning and explainability. An explanation-augmented dataset is constructed by leveraging a strong MLLM to generate explanations with gold sentiment constraints. Experimental results on Twitter2015 and Twitter2017 show improvements in both sentiment classification accuracy and explanation quality, demonstrating the approach's robustness and practical impact for explainable MABSA.

Abstract

Multimodal aspect-based sentiment analysis (MABSA) aims to identify aspect-level sentiments by jointly modeling textual and visual information, which is essential for fine-grained opinion understanding in social media. Existing approaches mainly rely on discriminative classification with complex multimodal fusion, yet lacking explicit sentiment explainability. In this paper, we reformulate MABSA as a generative and explainable task, proposing a unified framework that simultaneously predicts aspect-level sentiment and generates natural language explanations. Based on multimodal large language models (MLLMs), our approach employs a prompt-based generative paradigm, jointly producing sentiment and explanation. To further enhance aspect-oriented reasoning capabilities, we propose a dependency-syntax-guided sentiment cue strategy. This strategy prunes and textualizes the aspect-centered dependency syntax tree, guiding the model to distinguish different sentiment aspects and enhancing its explainability. To enable explainability, we use MLLMs to construct new datasets with sentiment explanations to fine-tune. Experiments show that our approach not only achieves consistent gains in sentiment classification accuracy, but also produces faithful, aspect-grounded explanations.

Explainable Multimodal Aspect-Based Sentiment Analysis with Dependency-guided Large Language Model

TL;DR

The paper reframes multimodal aspect-based sentiment analysis as a generative task that jointly outputs sentiment labels and natural language explanations using multimodal large language models. It introduces a dependency-syntax guided cue strategy to prune aspect-centered dependencies and textualizes them as prompts, enhancing aspect-aware reasoning and explainability. An explanation-augmented dataset is constructed by leveraging a strong MLLM to generate explanations with gold sentiment constraints. Experimental results on Twitter2015 and Twitter2017 show improvements in both sentiment classification accuracy and explanation quality, demonstrating the approach's robustness and practical impact for explainable MABSA.

Abstract

Multimodal aspect-based sentiment analysis (MABSA) aims to identify aspect-level sentiments by jointly modeling textual and visual information, which is essential for fine-grained opinion understanding in social media. Existing approaches mainly rely on discriminative classification with complex multimodal fusion, yet lacking explicit sentiment explainability. In this paper, we reformulate MABSA as a generative and explainable task, proposing a unified framework that simultaneously predicts aspect-level sentiment and generates natural language explanations. Based on multimodal large language models (MLLMs), our approach employs a prompt-based generative paradigm, jointly producing sentiment and explanation. To further enhance aspect-oriented reasoning capabilities, we propose a dependency-syntax-guided sentiment cue strategy. This strategy prunes and textualizes the aspect-centered dependency syntax tree, guiding the model to distinguish different sentiment aspects and enhancing its explainability. To enable explainability, we use MLLMs to construct new datasets with sentiment explanations to fine-tune. Experiments show that our approach not only achieves consistent gains in sentiment classification accuracy, but also produces faithful, aspect-grounded explanations.
Paper Structure (21 sections, 5 equations, 3 figures, 11 tables)

This paper contains 21 sections, 5 equations, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. The Approach
  3. Dependency Structure Extraction
  4. Aspect-centered Dependency Pruning
  5. MABSA with Dependencies
  6. Experiment Settings
  7. Dataset Construction for Explainable MABSA
  8. Baselines
  9. Implementation Details
  10. Evaluation Metrics
  11. Results and Analyses
  12. Overall Results
  13. Zero-shot Results
  14. Ablation Study
  15. Comparison of Textualization Approaches
  16. ...and 6 more sections

Figures (3)

  • Figure 1: The overall architecture of our approach (the pruned depth shown is three). The process includes three parts: the dependency syntax tree construction phase, the aspect-centered dependency pruning phase, the final textualization and MABSA with dependencies.
  • Figure 2: The number of samples (100 in total) that are rated as the best explanation by Qwen3-VL-32B. Blue, green, and yellow represent results of experimental groups: Vanilla, Syn ($n=\infty$), and Syn ($n=2$).
  • Figure 3: A presentation of sentiment classification results and partial explanations for the case. Our approach successfully distinguished the different sentiments conveyed by different aspects and generated correct explanations.