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

Aspect-Based Sentiment Intensity Analysis (ABSIA) has garnered increasing attention, though research largely focuses on domain-specific, sentence-level settings. In contrast, document-level ABSIA–particularly in addressing complex tasks like extracting Aspect-Category-Opinion-Sentiment-Intensity (ACOSI) tuples–remains underexplored. In this work, we introduce DanceHA, a multi-agent framework designed for open-ended, document-level ABSIA with informal writing styles. DanceHA has two main components: Dance, which employs a divide-and-conquer strategy to decompose the long-context ABSIA task into smaller, manageable sub-tasks for collaboration among specialized agents; and HA, Human-AI collaboration for annotation. We release Inf-ABSIA, a multi-domain document-level ABSIA dataset featuring fine-grained and high-accuracy labels from DanceHA. Extensive experiments demonstrate the effectiveness of our agentic framework and show that the multi-agent knowledge in DanceHA can be effectively transferred into student models. Our results highlight the importance of the overlooked informal styles in ABSIA, as they often intensify opinions tied to specific aspects.

💡 Research Summary

The paper introduces DanceHA, a novel multi‑agent framework designed to tackle the underexplored problem of document‑level Aspect‑Based Sentiment Intensity Analysis (ABSIA). While most prior work focuses on sentence‑level ABSA or ABSIA, the authors address the challenges of extracting fine‑grained Aspect‑Category‑Opinion‑Sentiment‑Intensity (ACOSI) tuples from long, informal user‑generated reviews. DanceHA consists of two main components: (1) “Dance,” a divide‑and‑conquer system that first splits a document into aspect‑based thought groups, then employs three specialized large language model (LLM) agents to assign a domain‑specific category, extract opinion expressions (preserving informal cues such as lengthening and emojis), and determine sentiment polarity together with a 0‑5 intensity score; the outputs are merged by a rule‑based integrator into structured ACOSI tuples. (2) “HA,” a Human‑AI collaboration layer where a Manager Agent aggregates the outputs of multiple Dance teams (built on different backbone LLMs such as DeepSeek‑V3 and GPT‑4o), resolves conflicts, and passes candidate annotations to human annotators for final verification and correction. This pipeline yields high‑quality labels with relatively low manual effort.

Using DanceHA, the authors construct Inf‑ABSIA, a new multi‑domain dataset comprising 2,714 long‑form reviews (≈90 words each) from laptop, restaurant, and hotel domains. Each document contains on average 8.48 ACOSI tuples, totaling 23,024 tuples—substantially richer than existing ABSA/ABSIA corpora. The dataset was created by first filtering for informal language (e.g., lengthened words) using a regex‑based detector, then applying the Dance pipeline, and finally validating the results through human review.

Extensive experiments evaluate seven LLMs across the three domains. DanceHA consistently outperforms few‑shot chain‑of‑thought (CoT) baselines, especially on texts with informal expressions where intensity prediction improves markedly. Moreover, the authors explore knowledge distillation: they extract reasoning chains from the Dance process and fine‑tune a smaller model (Qwen‑14B). The distilled model surpasses GPT‑4o few‑shot performance on all domains and approaches the results of a much larger model (Qwen2.5‑72B), demonstrating that the multi‑agent knowledge can be transferred efficiently.

Key contributions include: (i) a training‑free, modular multi‑agent architecture that decomposes a complex long‑context task into manageable sub‑tasks; (ii) the Inf‑ABSIA dataset, the largest publicly available document‑level ABSIA resource with explicit informal style annotations; (iii) empirical evidence that informal linguistic cues amplify sentiment intensity and should be explicitly modeled; and (iv) a successful distillation strategy that leverages the reasoning capabilities of the agentic system to improve smaller models.

Overall, DanceHA offers a practical solution for open‑ended, document‑level sentiment intensity extraction, highlights the importance of informal language in sentiment analysis, and provides a valuable benchmark for future research in long‑form opinion mining and multi‑agent LLM collaboration.