pith. sign in

arxiv: 2603.28205 · v2 · pith:MTQBGLQEnew · submitted 2026-03-30 · 💻 cs.CL

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

Yijin Wang , Fandi Sun , Haoyu Wen This is my paper

Pith reviewed 2026-05-15 06:38 UTC · model grok-4.3

classification 💻 cs.CL
keywords aspect-based sentiment analysiscomplex semantic spacesentiment polarityresidual projectionanti-collision maskABSAembedding entanglement
0
0 comments X

The pith

Projecting ABSA features into complex space via zero-initialized residuals separates opposing polarities by phase.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Aspect-based sentiment analysis struggles with entangled representations in standard embeddings, leading to poor separation of positive and negative opinions on aspects. This paper introduces a projection into complex space using zero-initialized residuals, where the phase distinguishes sentiment directions and the magnitude keeps related aspects aligned. An added anti-collision mask prevents same-polarity points from overlapping while pushing apart different polarities. On the ASAP dataset this yields a leading Macro-F1 of 0.8923, showing that complex representations can sharpen sentiment distinctions.

Core claim

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.

What carries the argument

Zero-Initialized Residual Complex Projection (ZRCP) that maps features to complex space for phase-based polarity isolation and amplitude regularization, paired with Anti-collision Masked Angle Loss to expand inter-polarity margins.

Load-bearing premise

Mapping features to complex space via zero-initialized residual projection plus the masked angle loss will reliably isolate polarities and generalize beyond the ASAP dataset without introducing new representation artifacts or requiring extensive hyperparameter tuning.

What would settle it

Running the model on a held-out ABSA benchmark and finding that the Macro-F1 score does not exceed that of standard real-valued models, or detecting increased collisions in the projected space.

Figures

Figures reproduced from arXiv: 2603.28205 by Fandi Sun, Haoyu Wen, Yijin Wang.

Figure 1
Figure 1. Figure 1: The overall architecture of our Phase-Driven Disentanglement framework. Textual inputs are encoded by a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Gradient Analysis of Fine-grained Contrastive Learning. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Model performance relative to RoBERTa across 18 fine-grained aspects. Bars represent the Macro-F1 score [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Similarity Matrix Analysis. (a) The baseline model without masks suffers from false negative collisions, resulting in a chaotic semantic space. (b) Our ZRCP+Mask framework delineates clear block-diagonal structures. (c) Model comparison on aggregated similarities shows that our model safely expands the inter-polarity discriminative margin (−7.3%) with only a marginal trade-off in intra-polarity cohesion. A… view at source ↗
Figure 5
Figure 5. Figure 5: Deep geometric analysis of complex amplitudes demonstrating the decoupling of subjective intensity from [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes a framework for Aspect-Based Sentiment Analysis using Zero-Initialized Residual Complex Projection (ZRCP) to embed textual features in complex space (with phase isolating polarities and amplitude regularized for consistency) together with an Anti-collision Masked Angle Loss that expands margins between opposing polarities while preserving intra-polarity cohesion. It reports a state-of-the-art Macro-F1 of 0.8923 on the ASAP dataset, outperforming baselines.

Significance. If the performance gains are shown to stem specifically from the complex-space phase separation rather than from the loss formulation or training choices alone, the approach could provide a useful alternative to real-valued embeddings and cosine similarity for handling representation entanglement in ABSA.

major comments (3)
  1. [Abstract] Abstract: the SOTA claim of Macro-F1 0.8923 is stated without any baseline names, training details, number of runs, or error bars, so the superiority cannot be assessed against the paper's own data.
  2. [Experiments] No ablation is reported that retains the masked angle loss but replaces the ZRCP complex projection with an equivalent real-valued residual; without this, it is impossible to attribute gains to phase-based polarity isolation rather than to the loss or schedule.
  3. [Method] No phase/amplitude histograms, angular-distance statistics, or t-SNE plots on the ASAP test set are supplied to confirm that the learned phase actually separates opposing polarities as claimed.
minor comments (1)
  1. [Abstract] The abstract refers to 'robust baselines' without naming them or citing their original papers.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript to strengthen the presentation and evidence.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the SOTA claim of Macro-F1 0.8923 is stated without any baseline names, training details, number of runs, or error bars, so the superiority cannot be assessed against the paper's own data.

    Authors: We agree that the abstract would benefit from additional context. In the revised version we will name the primary baselines, state that results are averaged over five runs, and include mean Macro-F1 with standard deviation. revision: yes

  2. Referee: [Experiments] No ablation is reported that retains the masked angle loss but replaces the ZRCP complex projection with an equivalent real-valued residual; without this, it is impossible to attribute gains to phase-based polarity isolation rather than to the loss or schedule.

    Authors: We acknowledge that an ablation isolating the complex projection is necessary. We will add this experiment, keeping the masked angle loss unchanged while substituting ZRCP with a real-valued residual projection of matching capacity, and report the resulting performance. revision: yes

  3. Referee: [Method] No phase/amplitude histograms, angular-distance statistics, or t-SNE plots on the ASAP test set are supplied to confirm that the learned phase actually separates opposing polarities as claimed.

    Authors: We agree that direct evidence for phase separation would strengthen the claims. We will include phase and amplitude histograms, angular-distance statistics between polarity classes, and t-SNE visualizations computed on the ASAP test set in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: novel projection and loss are independent proposals validated empirically

full rationale

The paper proposes ZRCP for mapping to complex space (phase isolating polarities, amplitude regularized) plus anti-collision masked angle loss as new components. No equations appear that reduce any claimed prediction or result to a fitted parameter by construction, nor does any load-bearing step collapse to a self-citation chain or imported uniqueness theorem. The Macro-F1 of 0.8923 is reported as an experimental outcome on ASAP, not derived from prior fits or renamed known patterns. The framework is self-contained against external benchmarks; performance gains are attributed to the described architecture rather than tautological re-use of inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

The abstract provides no explicit free parameters, axioms, or invented entities beyond naming the new projection and loss; all supporting assumptions about complex space benefits and dataset representativeness remain implicit and unstated.

invented entities (2)
  • Zero-Initialized Residual Complex Projection (ZRCP) no independent evidence
    purpose: Project textual features into complex semantic space to isolate sentiment polarities via phase
    Newly introduced method whose internal parameters and initialization are not detailed.
  • Anti-collision Masked Angle Loss no independent evidence
    purpose: Preserve intra-polarity cohesion while expanding margin between opposing polarities
    New loss function introduced to address false-negative collisions.

pith-pipeline@v0.9.0 · 5424 in / 1332 out tokens · 43487 ms · 2026-05-15T06:38:29.810879+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Issues and challenges of aspect-based sentiment analysis: A comprehensive survey

    Ambreen Nazir et al. “Issues and challenges of aspect-based sentiment analysis: A comprehensive survey”. In: IEEE Transactions on Affective Computing13.2 (2020), pp. 845–863

    work page 2020

  2. [2]

    Survey on aspect-level sentiment analysis

    Kim Schouten and Flavius Frasincar. “Survey on aspect-level sentiment analysis”. In:IEEE transactions on knowledge and data engineering28.3 (2015), pp. 813–830

    work page 2015

  3. [3]

    Contrastive sentence representation learning with adaptive false negative cancellation

    Lingling Xu et al. “Contrastive sentence representation learning with adaptive false negative cancellation”. In: Information Fusion102 (2024), p. 102065

    work page 2024

  4. [4]

    QPEN: quantum projection and quantum entanglement enhanced network for cross-lingual aspect-based sentiment analysis

    Xingqiang Zhao, Hai Wan, and Kunxun Qi. “QPEN: quantum projection and quantum entanglement enhanced network for cross-lingual aspect-based sentiment analysis”. In:Proceedings of the AAAI Conference on Artificial Intelligence. V ol. 38. 17. 2024, pp. 19670–19678. 11 APREPRINT- MARCH31, 2026

    work page 2024

  5. [5]

    Encoding word order in complex embeddings

    Benyou Wang et al. “Encoding word order in complex embeddings”. In:arXiv preprint arXiv:1912.12333(2019)

    work page arXiv 1912

  6. [6]

    A survey on aspect-based sentiment analysis: Tasks, methods, and challenges

    Wen Zhang et al. “A survey on aspect-based sentiment analysis: Tasks, methods, and challenges”. In:IEEE Transactions on Knowledge and Data Engineering35.11 (2022), pp. 11019–11038

    work page 2022

  7. [7]

    A syntactic approach for aspect based opinion mining

    T. C. Chinsha and S. Joseph. “A syntactic approach for aspect based opinion mining”. In:Proceedings of the 2015 IEEE 9th International Conference on Semantic Computing (IEEE ICSC 2015). IEEE, 2015, pp. 24–31

    work page 2015

  8. [8]

    Aspect based sentiment analysis using deep learning approaches: A survey

    G. S. Chauhan, R. Nahta, Y . K. Meena, et al. “Aspect based sentiment analysis using deep learning approaches: A survey”. In:Computer Science Review49 (2023), p. 100576

    work page 2023

  9. [9]

    BERT post-training for review reading comprehension and aspect-based sentiment analysis

    Hu Xu et al. “BERT post-training for review reading comprehension and aspect-based sentiment analysis”. In:Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). Association for Computational Linguistics, 2019, pp. 2324–2335

    work page 2019

  10. [10]

    Aspect based sentiment & emotion analysis with ROBERTa, LSTM

    U. Sirisha and S. C. Bolem. “Aspect based sentiment & emotion analysis with ROBERTa, LSTM”. In:Interna- tional Journal of Advanced Computer Science and Applications13.11 (2022)

    work page 2022

  11. [11]

    SimCSE: Simple contrastive learning of sentence embeddings

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. “SimCSE: Simple contrastive learning of sentence embeddings”. In:Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 2021, pp. 6894–6910

    work page 2021

  12. [12]

    Consistent positive correlation sample distribution: Alleviating the negative sample noise issue in contrastive adaptation

    Xing Wei et al. “Consistent positive correlation sample distribution: Alleviating the negative sample noise issue in contrastive adaptation”. In:Expert Systems with Applications263 (2025), p. 125669

    work page 2025

  13. [13]

    Complex-valued neural networks: A comprehensive survey

    Chiyan Lee, Hideyuki Hasegawa, and Shangce Gao. “Complex-valued neural networks: A comprehensive survey”. In:IEEE/CAA Journal of Automatica Sinica9.8 (2022), pp. 1406–1426

    work page 2022

  14. [14]

    Quantum-inspired neural language representation, matching and understanding

    Peng Zhang et al. “Quantum-inspired neural language representation, matching and understanding”. In:Founda- tions and Trends in Information Retrieval16.4-5 (2023), pp. 318–509

    work page 2023

  15. [15]

    Xianming Li and Jing Li.AnglE-optimized Text Embeddings. 2024. arXiv: 2309.12871 [cs.CL].URL: https: //arxiv.org/abs/2309.12871

    work page arXiv 2024

  16. [16]

    Quantum-inspired complex-valued language models for aspect- based sentiment classification

    Qin Zhao, Chenguang Hou, and Ruifeng Xu. “Quantum-inspired complex-valued language models for aspect- based sentiment classification”. In:Entropy24.5 (2022), p. 621

    work page 2022

  17. [17]

    Complex-Valued Vectors for Word Representation

    Guangcheng Liu and Yuexian Hou. “Complex-Valued Vectors for Word Representation”. In:2021 4th Interna- tional Conference on Artificial Intelligence and Big Data (ICAIBD). IEEE. 2021, pp. 242–246

    work page 2021

  18. [18]

    SDRS: Sentiment-aware disentangled representation shifting for multimodal sentiment analysis

    Sicheng Zhao et al. “SDRS: Sentiment-aware disentangled representation shifting for multimodal sentiment analysis”. In:IEEE Transactions on Affective Computing16.3 (2025), pp. 1802–1813

    work page 2025

  19. [19]

    Disentangled representation learning

    Xin Wang et al. “Disentangled representation learning”. In:IEEE Transactions on Pattern Analysis and Machine Intelligence46.12 (2024), pp. 9677–9696

    work page 2024

  20. [20]

    Exploiting adaptive contextual masking for aspect-based sentiment analysis

    SM Rafiuddin et al. “Exploiting adaptive contextual masking for aspect-based sentiment analysis”. In:Pacific- Asia Conference on Knowledge Discovery and Data Mining. Springer. 2024, pp. 147–159

    work page 2024

  21. [21]

    Enhancing aspect-based sentiment analysis with supervised contrastive learning

    Bin Liang et al. “Enhancing aspect-based sentiment analysis with supervised contrastive learning”. In:Pro- ceedings of the 30th ACM international conference on information & knowledge management. 2021, pp. 3242– 3247

    work page 2021

  22. [22]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Yinhan Liu et al. “RoBERTa: A Robustly Optimized BERT Pretraining Approach”. In: (2019). 12

    work page 2019