MASTE: A Multi-Agent Pipeline for Zero-Shot Aspect Sentiment Triplet Extraction

2026-07-09 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

MASTE, a multi-agent pipeline, significantly improves zero-shot Aspect Sentiment Triplet Extraction (ASTE) by addressing limitations of large language models (LLMs) in this task. Traditional LLMs struggle with ASTE's complexity, requiring them to determine span boundaries, opinion grouping, and sentiment polarity in a single decoding step, even with few-shot or chain-of-thought prompting. MASTE overcomes this by decomposing ASTE into four sequential stages, where specialized agents handle distinct compositional subtasks, explicitly conditioning on previous outputs. This design facilitates entirely training-free zero-shot ASTE, demonstrating strong generalization across various backbones and datasets. Extensive experiments on four ASTE benchmarks confirm MASTE's substantial outperformance over existing zero-shot and chain-of-thought LLM baselines, effectively narrowing the performance gap to fully supervised methods without requiring any labeled training data.

Key takeaway

For NLP Engineers developing zero-shot solutions for complex information extraction, MASTE's multi-agent pipeline offers a compelling alternative to single-pass LLM generation. You should consider decomposing intricate tasks like Aspect Sentiment Triplet Extraction into sequential, agent-handled subtasks. This approach can substantially improve performance without relying on labeled training data or extensive prompt engineering, narrowing the gap to supervised methods and streamlining deployment for new domains.

Key insights

Decomposing complex NLP tasks into multi-agent pipelines significantly enhances zero-shot LLM performance.

Principles

• Decompose complex tasks into sequential subtasks.
• Use specialized agents for each subtask.
• Condition agents on prior outputs.

Method

MASTE decomposes ASTE into four sequential stages, with specialized agents handling distinct compositional subtasks, explicitly conditioning on prior outputs for training-free zero-shot performance.

In practice

• Apply multi-agent design to complex NLP tasks.
• Improve zero-shot LLM performance without labeled data.
• Explore sequential subtask decomposition.

Topics

• Aspect Sentiment Triplet Extraction
• Zero-Shot Learning
• Multi-Agent Systems
• Large Language Models
• Natural Language Processing
• Information Extraction

Code references

• Hankerlove/MASTE

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis — Granularity-stratified multi-agent committees with divergence-aware routing enable cost-efficient, accurate financial sentiment analysis and hallucination detection.
• Sentiment analysis for software engineering: How far can zero-shot learning (ZSL) go? — Zero-shot learning offers a viable solution for sentiment analysis in software engineering, mitigating annotated data scarcity.
• AspectRAG: An Architecture of Retrieval and Generation for Aspect-Based Sentiment Analysis — AspectRAG leverages aspect-guided retrieval and generation for unsupervised, high-performance sentiment analysis in Portuguese.
• NCL HKU-NarrSim at SemEval-2026 Task 4: Aspect-Based Agents and Supervised Contrastive Embeddings for Narrative Similarity — Aspect-based agent frameworks and supervised contrastive learning effectively model narrative similarity beyond surface lexical cues.
• Extracting Problem and Method Sentence from Scientific Papers: A Context-enhanced Transformer Using Formulaic Expression Desensitization — Formulaic expression desensitization and context-enhanced transformers improve problem/method sentence extraction from scientific papers, outperforming baselines.
• LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining — Constraining LLM output space via retrieval and validation significantly improves structured prediction accuracy and robustness.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.