Research map / Multi-agent systems and games / LLM agents

DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DanceHA, a training-free multi-agent framework (Dance divide-and-conquer teams plus Human-AI annotation) that builds the Inf-ABSIA document-level ACOSI dataset from informal reviews and distills its reasoning into Qwen3 students.

Problem

Document-level aspect-based sentiment intensity analysis (ACOSI tuple extraction) on informal user-generated reviews is underexplored, mainly due to the lack of fine-grained labeled datasets.

Approach

Decompose long-document ABSIA into aspect-based thought groups handled by specialized LLM agents (category, opinion, sentiment/intensity) with rule-based merging, then use multiple Dance teams aggregated by Manager Agents and human revision to create labels; distill the resulting reasoning chains into small models via LoRA.

Key results

• MA with DeepSeek-V3 reaches 63.18% F1 on Restaurant versus 47.80% for the best single Dance team (GPT-4o).
• Few-shot CoT GPT-4o reaches 35.04% F1 on Restaurant, below Dance with GPT-4o.
• SFT Qwen3-14B reaches 34.52% F1 on Restaurant, outperforming 5-shot CoT GPT-4o (35.04 is higher, so comparison is on other domains; reported as outperforming across domains in text).

Surprising findings

• Qwen3-4B with zero-shot thinking-mode CoT outperforms non-thinking Dance, while for Qwen3-14B Dance matches or beats thinking CoT.
• Sentiment intensity prediction is the hardest sub-task, with average accuracy around 61%, despite polarity accuracy above 94%.

Limitations

• Influence of team size on annotation quality is unexplored.
• Labels are generated by LLM agents with moderate human revision; inter-annotator agreement is only 0.61.
• Sentiment intensity prediction and divider sub-tasks remain difficult for LLMs.

Experimental setup

Datasets: Amazon Reviews-Electronics Yelp TripAdvisor

Benchmarks: Inf-ABSIA

Metrics: F1 accuracy MAE polarity accuracy SIS accuracy

Baselines: zero-shot CoT 5-shot CoT GPT-4o GPT-4o-mini GPT-3.5-turbo DeepSeek-V3 Qwen2.5-72B-Instruct Qwen3-4B Qwen3-14B

Builds on: chain-of-thought prompting batch prompting retrieval-augmented generation

Closest papers by meaning

• Aspect-Based Sentiment Analysis with Explicit Sentiment Augmentations (AAAI 2024 · similarity 0.89)
• Intelligent Calibration for Bias Reduction in Sentiment Corpora Annotation Process (AAAI 2024 · similarity 0.84)
• You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.84)
• Aspect Enhancement and Text Simplification in Multimodal Aspect-Based Sentiment Analysis for Multi-Aspect and Multi-Sentiment Scenarios (AAAI 2025 · similarity 0.84)
• A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.83)
• TextGT: A Double-View Graph Transformer on Text for Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.83)