LLM-MvP: An Efficient Solution for Aspect-Based Sentiment Analysis Based on Multi-View Prompting

LLM-MvP is an aspect-based sentiment analysis (ABSA) solution based on multi-view prompting. Through three key technologies—multi-view prompting, schema-constrained decoding, and prefix batching—it achieves or even surpasses the performance of fine-tuned models under few-shot conditions, while significantly reducing computational overhead. Original Authors and Source: The paper's author team (submitted to arXiv), original title: Prompting Is All You Need: Multi-view Prompting Large Language Models for Aspect-Based Sentiment Analysis, link: http://arxiv.org/abs/2605.28058v1, published on May 27, 2026.

Aspect-based sentiment analysis (ABSA) is a fine-grained NLP task that requires identifying the sentiment orientation of specific aspects in text (e.g., in "The food at this restaurant is delicious, but the service is too slow", "food" is positive and "service" is negative). Traditional ABSA relies on large amounts of labeled data for supervised fine-tuning; while few-shot prompting with LLMs has made progress, it has two major issues: 1. Performance gap compared to fine-tuned models with hundreds of samples; 2. High inference computational overhead, limiting deployment.

Core Idea: Derived from multi-view learning, it generates multiple view prompts to allow the model to examine text from multiple angles and comprehensively capture aspect sentiment information. Three Core Components:

1. Schema-constrained decoding: Uses context-free grammar (CFG) to constrain the output format, ensuring compliance with ABSA's structural requirements;
2. Prefix batching: Shares the computation of common prefixes across different view prompts, reducing redundant overhead and processing multiple views in a single forward pass;
3. Multi-view aggregation: A lightweight mechanism that synthesizes confidence from each view to output the final judgment.

Evaluated on five standard ABSA benchmark datasets (covering domains like restaurant and laptop reviews), key findings:

1. Performance on par with fine-tuning: Achieves or surpasses specialized fine-tuned models with hundreds of samples across multiple datasets;
2. Significant computational efficiency: Prefix batching drastically reduces the inference overhead of multi-view processing;
3. Few-shot advantage: Achieves high performance with only a few examples, reducing reliance on large-scale labeled data.

Practical Value: Provides an efficient ABSA solution that can be deployed without large-scale labeling or expensive fine-tuning. Application Scenarios:

• Resource-constrained scenarios: Achieve high-performance sentiment analysis at low cost;
• Rapid iteration needs: Prompt engineering is more flexible than fine-tuning, allowing quick adjustments to analysis dimensions;
• Multilingual expansion: Leverage the multilingual capabilities of LLMs to support cross-lingual ABSA.

The success of LLM-MvP reveals trends:

1. The upper limit of prompt engineering exceeds expectations: Structured multi-view design can match the performance of traditional fine-tuning;
2. Decoding strategies are equally important: Schema-constrained decoding improves task performance by controlling generation;
3. Computational efficiency is core to deployment: Technologies like prefix batching drive LLMs from research to production.