MASTE multi-agent pipeline enables training-free zero-shot aspect sentiment triplet extraction

Researchers propose MASTE, a multi-agent pipeline designed to improve zero-shot Aspect Sentiment Triplet Extraction (ASTE) by decomposing the task into four sequential stages handled by specialized agents. This approach addresses the limitations of single-pass generation in large language models, which struggle to determine span boundaries, opinion grouping, and sentiment polarity simultaneously.

• MASTE eliminates the need for labeled training data or few-shot demonstrations by using explicit conditioning on prior outputs across its stages.
• The framework generalizes across different model backbones and datasets while outperforming zero-shot and chain-of-thought baselines.
• Extensive experiments on four ASTE benchmarks demonstrate that MASTE narrows the performance gap to fully supervised methods without using any labeled triplets.

The authors consider this significant because it provides a broadly available, training-free solution for complex NLP tasks where in-domain data is typically unavailable.