Beyond Cosine Similarity: Zero-Initialized Residual Complex Projection for Aspect-Based Sentiment Analysis

Yijin Wang; Fandi Sun; Haoyu Wen

cs.CLMar 30, 2026

Beyond Cosine Similarity: Zero-Initialized Residual Complex Projection for Aspect-Based Sentiment Analysis

Authors: Yijin Wang, Fandi Sun, Haoyu Wen

Organizations: School of Economics and Management, Xidian University, Xi’an, China

Abstract

Aspect-Based Sentiment Analysis (ABSA) faces critical challenges due to representation entanglement and false-negative collisions in real-valued embedding spaces. In this paper, we propose a novel framework featuring a Zero-Initialized Residual Complex Projection (ZRCP) and an Anti-collision Masked Angle Loss. Our approach projects textual features into a complex semantic space, utilizing the phase to isolate sentiment polarities while regularizing the amplitude to ensure structural consistency within aspect categories. To mitigate this, we introduce an anti-collision mask that preserves intra-polarity aspect cohesion while significantly expanding the discriminative margin between opposing polarities. Experimental results on the ASAP dataset demonstrate that our framework achieves a state-of-the-art Macro-F1 score of 0.8923, outperforming robust baselines.

Figures & tables

Figure 1: Overall architecture of our Phase-Driven Disentanglement framework. Textual inputs are mapped into a complex space via the ZRCP module and optimized using an Anti-collision Masked Angle Loss to decouple aspect and sentiment.

Figure 2: Theoretical comparison of Cosine and Angle Loss. (a) In the dense pre-trained initialization zone ( $\theta\to 0^{\circ}$ ), the cosine objective saturates into a flat landscape. (b) This saturation causes severe vanishing gradients ( $|\sin\theta|\to 0$ ). In contrast, our angle-based objective maintains a constant driving force to effectively separate entangled hard negatives.

Aspect	Zero-shot		RoB-pair		LCFS		DualGCN		SimCSE		AnglE (Form C)		Ours (AnglE+ZRCP)
Aspect	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc
Transportation	0.5359	0.6950	0.6559	0.8571	0.6834	0.8861	0.7950	0.9323	0.9039	0.9710	0.8622	0.9592	0.9022	0.9721
Downtown	0.4706	0.7217	0.5700	0.9245	0.6084	0.9363	0.7454	0.9658	0.7961	0.9682	0.8434	0.9776	0.7858	0.9717
Easy to find	0.5998	0.6601	0.7150	0.7696	0.7534	0.8076	0.9004	0.9297	0.9191	0.9424	0.9149	0.9401	0.9265	0.9482
Queue	0.6830	0.6851	0.7502	0.7511	0.7649	0.7660	0.7802	0.7809	0.8401	0.8404	0.8573	0.8574	0.8743	0.8745
Hospitality	0.7932	0.8526	0.8678	0.9039	0.8793	0.9138	0.9281	0.9497	0.9505	0.9660	0.9479	0.9642	0.9582	0.9714

Table 1: Performance comparison on the ASAP dataset. Best results are bolded ; second-best are underlined .

Ablation on Network & Objective
Model Variations	Macro-F1	Acc
Ours (Full Framework)	0.8923	0.9418
$w/o$ ZRCP (Hard Chunking)	0.8739	0.9319
$w/o$ Anti-collision Mask	0.8809	0.9336
$w/o$ Angle Loss ( $w_{angle}=0$ )	0.8843	0.9389
Ablation on Data Strategy

Table 2: Ablation study of architectural components and data strategies.

Figure 3: Similarity matrix analysis. (a) The unmasked baseline suffers from false negative collisions. (b) Our method delineates clear block-diagonal structures. (c) Our model expands the inter-polarity discriminative margin ( $-6.5\%$ ) while preserving intra-polarity cohesion.

Model	SemEval-2016		MAMS-ACSA
Model	Macro-F1	Acc	Macro-F1	Acc
RoB-pair	0.8180	0.8869	0.6383	0.7193
DualGCN	0.8289	0.9135	62.643	66.548
SimCSE	0.8457	0.8962	0.6038	0.7217
AnglE (Format C)	0.8632	0.9134	–	–
AnglE (Format A)	0.8784	0.9078	0.6532	0.7526

Table 3: Overall performance comparison on SemEval-2016 and MAMS-ACSA datasets. Best results are bolded ; second-best are underlined .

Figure 4: Statistical verification of length-amplitude decoupling ( $r=0.130,p<0.001$ ), demonstrating robustness against physical text length biases.

Aspect Category	w/o AmpLoss	with AmpLoss	$\Delta$
Objective / Fact-based Aspects
Transportation	0.8916	0.9022	+1.06%
Downtown	0.7489	0.7858	+3.69%
Subjective / Descriptive Aspects
Decoration	0.8556	0.8627	+0.71%
Taste	0.9265	0.9247	-0.18%

Table 4: Impact of Amplitude Penalty ( AmpLoss ) on divergent aspects.

Aspect	RoB-pair		DualGCN		SimCSE		AnglE (Format C)		AnglE (Format A)		Ours (ZRCP+Mask)
Aspect	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc
Restaurant General	0.9549	0.9701	0.8923	0.9220	0.9780	0.9851	0.9780	0.9851	1.0000	1.0000	0.9780	0.9851
Service	1.0000	1.0000	0.9122	0.9122	0.9787	0.9825	0.9581	0.9649	0.9787	0.9825	0.9566	0.9649
Food Quality	0.9424	0.9737	0.8493	0.9433	0.8848	0.9474	0.9699	0.9868	0.9699	0.9868	0.9171	0.9605
Food Style/Options	0.7013	0.7826	0.8627	0.8913	0.7319	0.7826	0.7745	0.8261	0.7745	0.8261	0.7319	0.7826
Food Prices	0.8611	0.8667	0.8400	0.8500	0.9321	0.9333	0.9321	0.9333	1.0000	1.0000	0.9321	0.9333

Table 5: Detailed per-aspect performance on the English SemEval-2016 dataset. Best results are bolded ; second-best are underlined .

Aspect	RoB-pair		DualGCN		SimCSE		AnglE		Ours (AnglE+ZRCP)
Aspect	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc
Food	0.7298	0.7526	0.5237	0.5464	0.6008	0.6564	0.6920	0.7354	0.7224	0.7698
Service	0.7702	0.8077	0.5771	0.7308	0.6737	0.7308	0.7676	0.8205	0.8384	0.8590
Staff	0.6108	0.7870	0.5925	0.8284	0.6166	0.8047	0.5995	0.8047	0.6199	0.8047
Price	0.7540	0.7368	0.6735	0.6579	0.6647	0.6579	0.7868	0.7895	0.7210	0.7105
Ambience	0.5302	0.7188	0.5494	0.8125	0.5703	0.8438	0.5297	0.7188	0.5523	0.7500

Table 6: Detailed per-aspect performance on the MAMS-ACSA dataset. Best results are bolded and second-best are underlined .

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
SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)
Apr 8, 2026Liang-Chih Yu, Jonas Becker, Shamsuddeen Hassan Muhammad +14Sentiment Analysis Aspect-Based Sentiment Analysis
Yuan Ze University · University of Göttingen · Imperial College London +10

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

SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)