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

Lei Wang; Min Huang; Eduard Dragut

cs.CLMar 17, 2026

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

Authors: Lei Wang, Min Huang, Eduard Dragut

Organizations: Temple University, Philadelphia, PA, USA · Independent Researcher, Philadelphia, PA, USA

Abstract

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.

Figures & tables

Figure 1: Overview of DanceHA , consisting of two key components: (1) Dance ( D ivide- an d- C onquer T e amwork) for open-ended document-level ABSIA and (2) HA ( H uman- A I collaboration) for label annotation. Figure 2 illustrates Dance with an example.

Dataset	# D	Level	Topic	Lang	ACOS	Intensity	Informal	Size	Avg # w	Avg # Tuple	# Tuple
SemEval2016-SB1 ( Pontiki et al. 2016 )	2	Sentence	ABSA	E	Y	N	N	5,984	13	1.18	7,061
ACOS-Rest&Lap ( Cai et al. 2021 )	2	Sentence	ABSA	E	Y	N	N	6,362	16	1.48	9,415
ABSIA ( Mamta and Ekbal 2023 )	1	Sentence	ABSIA	E	N	Y	N	4,650	17	1.00	4,650
SIGHAN2024 ( Lee et al. 2024 )	1	Sentence	ABSIA	C	Y	Y	N	8,150	15	1.50	12,225
SemEval2016-SB2 ( Pontiki et al. 2016 )	2	Document	ABSA	E	N	N	N	900	60	4.96	4,464
DOTSA ( Luo et al. 2022 )	6	Document	ABSA	E	N	N	N	1,194	23	1.26	1,504

Table 1: Statistics comparison of Inf-ABSIA with existing ABSA/ABSIA datasets. Inf-ABSIA has larger sample size (2,714 vs. 1,194), longer context (avg. words 90 vs. 60), and more fine-grained labels (avg. tuples 8.48 vs. 4.96). # D indicates the number of domains. # w denotes the number of words. Lang denotes the language, where E stands for English and C for Chinese.

Figure 2: An example of Dance for document-level ABSIA.

Domain	Method	Model	F1 (%)	Acc (%)	MAE
Rest	zs-CoT	GPT-4o	22.95 $\pm$ 1.14	12.97 $\pm$ 0.73	0.467 $\pm$ 0.053
	fs-CoT	GPT-4o	35.04 $\pm$ 2.43	21.27 $\pm$ 1.80	0.428 $\pm$ 0.023
	Dance	GPT-3.5-turbo	4.89 $\pm$ 0.31	2.50 $\pm$ 0.16	0.791 $\pm$ 0.073
	Dance	GPT-4o-mini	19.90 $\pm$ 0.38	11.05 $\pm$ 0.23	0.418 $\pm$ 0.007
	Dance	GPT-4o	47.80 $\pm$ 1.65	31.42 $\pm$ 1.41	0.210 $\pm$ 0.019
	Dance	DeepSeek-V3	37.93 $\pm$ 1.75	23.42 $\pm$ 1.33	0.274 $\pm$ 0.016

Table 2: Performance comparison of Dance, MA, zero-shot (zs) and 5-shot (fs) CoT with various base models on Inf-ABSIA. F1 and Acc represent the evaluation metrics for the ACOS tuple extraction, MAE for SIS.

\pm

indicates std score. We use boldface to indicate the best performance within each method.

Expert Agents		Divider ( $F_{D}$ )			Category ( $F_{C}$ )	Opinion ( $F_{O}$ )	Sentiment Polarity & Intensity ( $F_{S}$ )
Sub-tasks		Thought Group	Aspect	All	All	All	Polarity	SIS	All
Rest	GPT-3.5-turbo	17.71 $\pm$ 0.16	8.54 $\pm$ 1.22	2.40 $\pm$ 0.26	75.33 $\pm$ 1.88	36.87 $\pm$ 3.42	96.69 $\pm$ 0.80	40.06 $\pm$ 2.76	38.73 $\pm$ 2.68
	GPT-4o-mini	20.32 $\pm$ 0.19	29.66 $\pm$ 1.48	10.29 $\pm$ 0.28	78.37 $\pm$ 1.29	46.77 $\pm$ 2.86	95.20 $\pm$ 1.98	59.52 $\pm$ 0.48	56.66 $\pm$ 1.22
	GPT-4o	54.11 $\pm$ 1.49	54.53 $\pm$ 1.73	36.01 $\pm$ 1.21	82.58 $\pm$ 0.69	68.39 $\pm$ 1.07	95.96 $\pm$ 0.49	73.43 $\pm$ 1.25	70.47 $\pm$ 1.54
	DeepSeek-V3	51.81 $\pm$ 1.09	46.47 $\pm$ 0.82	28.34 $\pm$ 0.35	81.79 $\pm$ 0.71	61.06 $\pm$ 0.93	95.14 $\pm$ 1.64	70.52 $\pm$ 2.18	67.09 $\pm$ 2.43
	Qwen3-4B	24.40 $\pm$ 0.17	26.90 $\pm$ 2.65	9.53 $\pm$ 0.76	72.77 $\pm$ 0.33	35.43 $\pm$ 2.34	96.59 $\pm$ 1.97	39.01 $\pm$ 2.88	37.69 $\pm$ 3.10

Table 3: Evaluation of Expert Agents in Dance and their sub-tasks. The evaluation metric is accuracy (%). We use boldface to indicate the best performance and underline to indicate the second-best.

Figure 3: SIS scores for ACOSI tuples with informal and formal styles among different baselines. Error bars indicate standard deviation score. Results demonstrate the influence of informal styles on sentiment intensity and the capability of advanced models to interpret informal expressions.

	Method	F1(%)	Acc(%)	MAE
Rest	Dance	35.38	21.50	0.364
	w/o D&C	27.19 $(8.19\downarrow)$	15.73 $(5.76\downarrow)$	0.473 $(0.109\uparrow)$
	w/o Teamwork	32.48 $(2.90\downarrow)$	19.40 $(2.10\downarrow)$	0.456 $(0.093\uparrow)$
Hotel	Dance	32.76	19.61	0.327
	w/o D&C	22.28 $(10.48\downarrow)$	12.54 $(7.07\downarrow)$	0.490 $(0.163\uparrow)$
	w/o Teamwork	29.81 $(2.95\downarrow)$	17.53 $(2.08\downarrow)$	0.479 $(0.151\uparrow)$

Table 4: Ablation results for the Dance component. w/o D&C removes the Divide-and-Conquer Strategy; w/o Teamwork uses a single agent for the Conquer phase. Red indicates performance drops.

	Method	Model	Mode	F1 (%)	Acc (%)	MAE
Rest	zs-CoT	4B	think	20.22 $\pm$ 2.34	11.27 $\pm$ 1.44	1.238 $\pm$ 0.068
	zs-CoT	14B	think	19.06 $\pm$ 2.80	10.56 $\pm$ 1.73	0.740 $\pm$ 0.100
	Dance	4B	none	13.68 $\pm$ 1.13	7.34 $\pm$ 0.65	0.909 $\pm$ 0.051
	Dance	14B	none	25.87 $\pm$ 1.85	14.87 $\pm$ 1.23	1.016 $\pm$ 0.010
	SFT	4B	think	30.82 $\pm$ 1.14	18.22 $\pm$ 0.80	0.408 $\pm$ 0.026
	SFT	14B	think	34.52 $\pm$ 1.23	20.87 $\pm$ 0.91	0.395 $\pm$ 0.031

Table 5: Results for distillation with reasoning chains. We use boldface to indicate the best performance and underline to indicate the second-best. Model 4B denotes Qwen3-4B; 14B denotes Qwen3-14B.

	Rest	Hotel	Laptop	Avg
Rater	4.62	4.55	4.25	4.47
IAA	0.65	0.61	0.58	0.61

Table 6: Human evaluation results for data quality.

Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) identifies what people express opinions about and the sentiment or emotion attached to each aspect.

See all 52 papers

DiffuSent: Towards a Unified Diffusion Framework for Aspect-Based Sentiment Analysis
May 31, 2026Shu Long, Yanglei Gan, Xuchuan ZhouAspect-Based Sentiment Analysis Conditional Diffusion Models
University of Electronic Science and Technology of China · Southwest Petroleum University · Southwest Minzu University
ACAT: A Collaborative Platform for Efficient Aspect-Based Sentiment Dataset Annotation
Jun 2, 2026Ana-Maria Luisa Mocanu, Ciprian-Octavian Truica, Elena-Simona ApostolHuman-in-the-Loop Annotation Aspect-Based Sentiment Analysis
1National University of Science and Technology POLITEHNICA Bucharest, Romania · 2Academy of Romanian Scientists, Ilfov 3, Bucharest, 050044, Romania
AnnoABSA: A Web-Based Annotation Tool for Aspect-Based Sentiment Analysis with Retrieval-Augmented Suggestions
Mar 2, 2026Nils Constantin Hellwig, Jakob Fehle, Udo Kruschwitz +1Aspect-Based Sentiment Analysis LLM-Assisted Annotation
Media Informatics Group, University of Regensburg, Regensburg, Germany · Information Science Group, University of Regensburg, Regensburg, Germany

May 31, 2026J/K move · Enter open · S save

DiffuSent: Towards a Unified Diffusion Framework for Aspect-Based Sentiment Analysis

Shu Long, Yanglei Gan, Xuchuan Zhou

University of Electronic Science and Technology of China · Southwest Petroleum University · Southwest Minzu University

Aspect-Based Sentiment Analysis Conditional Diffusion Models

Open paper AskPDF arXiv

Aspect-Based Sentiment Analysis (ABSA) encompasses seven distinct subtasks, each focusing on different extracted elements. Despite the proven success of generative models in unified aspect sentiment analysis, existing approaches often rely on auto-regressive token-by-token generation without grasping the whole information of the aspect and opinion terms, resulting in boundary insensitivity, particularly in context of multi-word aspect and opinion terms. To address these issues, we present DiffuSent, a non-auto-regressive diffusion framework that systematically formulates all ABSA subtasks as boundary denoising diffusion processes, progressively refining boundaries over noisy states. Furthermore, we introduce a contrastive denoising training strategy which effectively address duplicate predictions with subtle variations introduced by diffusion process. Extensive experiments across 28 settings (7 subtasks x 4 datasets) demonstrate that DiffuSent achieves delivers consistent improvements over the strongest generative and span-based systems. DiffuSent exhibits notable gains on multi-word triplets, achieving an average improvement of +2.48 F1, and maintains robust extraction accuracy in sentences containing multiple sentiment triplets. Moreover, the non-auto-regressive decoding enables substantial efficiency benefits, reaching up to 181 times faster inference than auto-regressive generative baselines

Abstract

Aspect-Based Sentiment Analysis

Explore similar work

DiffuSent: Towards a Unified Diffusion Framework for Aspect-Based Sentiment Analysis

ACAT: A Collaborative Platform for Efficient Aspect-Based Sentiment Dataset Annotation

AnnoABSA: A Web-Based Annotation Tool for Aspect-Based Sentiment Analysis with Retrieval-Augmented Suggestions