Pith. sign in

REVIEW 5 major objections 5 minor 49 references

AF-MAT: Aspect-aware Flip-and-Fuse xLSTM for Aspect-based Sentiment Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An xLSTM architecture tops aspect-sentiment accuracy on three benchmarks

desk verdict AF-MAT is a plausible new combination of xLSTM for ABSA, but the paper's central aspect gate as written is a no-op and the forward pass is under-specified, making the claimed SOTA gains unverifiable without code or a corrected derivation. read the letter →

arxiv 2507.01213 v2 pith:ZOLDWA3P submitted 2025-07-01 cs.CL

Adamu Lawan , Juhua Pu , Haruna Yunusa , Muhammad Lawan , Mahmoud Basi , Muhammad Adam This is my paper · ORCID
classification cs.CL
keywords aspect-basedsentimentanalysisxLSTMmaspectgateFlipMixmultiheadcross-featurefusionsequencereversallong-rangedependencies
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that an architecture built on xLSTM can beat existing methods for aspect-based sentiment analysis while staying efficient. It introduces AF-MAT, which combines a matrix-LSTM memory with an aspect gate, a flip-based local-and-long-range encoder, and a lightweight fusion block. On three standard benchmarks—SemEval 2014 Restaurant and Laptop and a Twitter dataset—the authors report the best accuracy among non-BERT models and, when combined with BERT, accuracy above all compared baselines (for example 87.72 on Restaurant14). If the result holds, it would give ABSA a linear-time alternative to attention and Mamba-based models that does not depend on CUDA-specific kernels.

What carries the argument

The machinery is three modules. AA-mLSTM injects an aspect gate $a_t$ into the cumulative forget/input decay $\log D_{t,j} = \log F_{t,j} + \log i_j + \log a_j$, so that memory updates are scaled by aspect relevance. FlipMix first partially reverses the sequence (keeping $r$ elements) and applies a partially flipped Conv1D to capture short-range patterns, then fully reverses and applies mLSTM for long-range patterns. MC2F fuses the two paths using an mLSTM where AA-mLSTM outputs serve as queries and keys and FlipMix outputs as values, giving a linear-time soft accumulation of context instead of one-step attention.

What would settle it

Compute $a_t$ from Equation (4) for each token in any sentence: the input to the exponential is the same broadcast vector at every $t$, so $a_t$ is constant across positions. A reader can verify that replacing $a_t$ with a single learnable scalar (or omitting it) leaves the forward pass unchanged up to a global scale, and if the reported accuracy does not drop, the aspect gate is not doing the token-level selective-emphasis work claimed.

Watch

Extended reading notes

Core claim

The central discovery is that the xLSTM memory update can be made aspect-aware by inserting a dedicated aspect gate into the stabilized decay matrix of an mLSTM, and that flipping parts of the sequence before convolution and fully reversing before mLSTM lets a single linear-time model capture both local and long-range aspect-sentiment dependencies. The authors show that this combination, with an mLSTM-gated multihead fusion between the forward and flipped paths, outperforms all compared baselines on three ABSA datasets, with AF-MAT+BERT reaching an accuracy of 87.72 on Restaurant14, 81.87 on Laptop14, and 78.54 on Twitter.

Load-bearing premise

Equation (4) defines the aspect gate as $\exp(W_a[\tilde{H}_a \oplus \tilde{H}_a \oplus \tilde{H}_a]+b_a)$, which depends only on the pooled aspect vector and is therefore identical for every token position, so the claimed per-token selective emphasis requires that this gate actually vary with the token being processed.

Editorial extensions

If this is right

  • AF-MAT is the first xLSTM-based model for ABSA and, according to the reported experiments, sets a new state of the art among non-BERT models on Restaurant14, Laptop14, and Twitter.
  • When combined with BERT, AF-MAT exceeds every compared BERT-based baseline on all three datasets, including the MambaForGCN+BERT prior best.
  • The partial-flip hyperparameter $r$ is consequential: too large or too small degrades accuracy, with $r=6$ optimal on the Restaurant dataset, indicating that the balance of local versus reversed distance matters.
  • Ablation results show that removing any of the aspect gate, pf-Conv1D, ff-mLSTM, or MC2F lowers accuracy on all three datasets, implying each component contributes to the final result.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the aspect gate in Equation (4) is really computed only from the pooled aspect vector, then it is constant across all token positions and the claimed per-token selective emphasis cannot be realized as written; a token-dependent gate, e.g. including the token's own hidden state inside the exponential, would be the natural correction and is probably what the authors intended.
  • The 'fully flipped' mLSTM path may partially cancel the directional cues introduced by the partial flip, so the optimal $r$ could be dataset-dependent; testing the reported $r=6$ optimum on Laptop and Twitter would clarify whether the mechanism generalizes.
  • Because MC2F is described as an mLSTM-gated fusion of two feature streams, the same design could transfer to other target-conditioned sequence tasks like stance detection or targeted emotion classification, where a target gate and reversed context pass could help similarly.
  • The paper leaves the definitions of the mLSTM's own $q_t$, $v_t$, $i_t$, $f_t$ to an unresolved 'Equation ??' reference; a complete specification of those terms is required for the reported numbers to be independently reproducible.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes AF-MAT, an xLSTM-based framework for aspect-based sentiment analysis (ABSA). The architecture combines an Aspect-aware mLSTM (AA-mLSTM) with a dedicated aspect gate, a FlipMix block that applies partially and fully flipped convolutions and mLSTM layers, and an MC2F fusion module. The authors evaluate AF-MAT and AF-MAT+BERT on Restaurant14, Laptop14, and Twitter, reporting accuracy and macro-F1 numbers that they interpret as state-of-the-art, and they include an ablation study, hyperparameter analyses for the flip length r and layer count L, and qualitative case studies.

Significance. If the proposed architecture worked as described, the paper would offer a useful new direction for ABSA by importing xLSTM and combining reversed multi-scale context modeling with aspect conditioning. The paper is among the first to apply xLSTM to ABSA, and the modular design is well motivated. It also gives credit to the efficiency goal of avoiding quadratic attention and CUDA-specific kernels. However, the central mechanism as written is not correct: the aspect gate in Eq. (4) is token-independent and cancels in the normalization, the forward pass is incomplete because the definitions of q, k, v, i, and f are missing, and the reported empirical comparison lacks variance estimates and appears to tune hyperparameters on the test split. These issues directly undermine the paper's main claims, so the manuscript is not close to acceptable in its current form. No code or data are provided, which further limits verifiability.

major comments (5)
  1. [Aspect-aware mLSTM (AA-mLSTM), Eq. (4)] Equation (4) defines a_t as a function only of the mean-pooled aspect representation ~H_a^fwd, with no token-dependent quantity on the right-hand side. Hence a_t is a per-sample constant c. Substituting this into Eq. (5) adds the same value log c to every log D_{t,j}; the max subtraction in Eq. (7) removes it, and even if it were retained, the normalization in Eq. (10) cancels a constant factor in C_{t,j}. The aspect gate therefore has exactly zero effect on the AA-mLSTM output as written. This contradicts the claim that the gate selectively emphasizes tokens relevant to the aspect, and it makes the Table 4 ablation self-contradictory: removing a no-op gate cannot reduce Restaurant14 accuracy by 1.44 points. The authors must provide a correct token-dependent gate formulation or explain how their implementation differs from Eq. (4).
  2. [Aspect-aware mLSTM (AA-mLSTM), Eqs. (8)-(10)] The forward pass in Eqs. (8)-(10) computes alpha_{t,j} = q_t^T k_j / sqrt(d) for every pair (t,j), applies a causal normalization over j, and forms H_t as a weighted sum of v_j. This is exactly causal dot-product attention and requires O(N^2) time in the sequence length, contradicting the paper's repeated claim of a linear-time framework. If the authors intend a recurrent mLSTM update rather than an attention-style pairwise computation, the equations must be rewritten to show that; as stated, the linear-time claim is not supported.
  3. [Aspect-aware mLSTM (AA-mLSTM), after Eq. (11)] The text says 'We compute the q_t, v_t, i_t, and f_t as described in Equation ??, using ~H^fwd as input,' but no such equation appears in the manuscript. Since q, k, v, i, and f are the core variables of the proposed mLSTM, the AA-mLSTM forward pass cannot be reconstructed or independently implemented from the paper. This missing specification must be supplied before the architecture claims can be assessed.
  4. [Effect of Hyperparameter r and Effect of Hyperparameter L] The hyperparameter analyses in Figures 3 and 4 select r=6 and L=2 by maximizing accuracy on the Restaurant dataset, and the same Restaurant results then appear in the main Table 2. If the test split was used for selection, the reported Restaurant14 accuracy is optimistically biased. The authors should tune on a held-out validation split or use nested cross-validation, and they should clarify how the selected r and L are transferred to Laptop14 and Twitter.
  5. [Main Results, Table 2] All results in Table 2 are single numbers with no standard deviations or statistical significance tests, and the strongest baselines are not re-implemented under the same training protocol. Several reported improvements are very small (e.g., 78.71 vs. 78.64 accuracy on Laptop14 for AF-MAT vs. MambaForGCN), so the state-of-the-art claim is not supported without variance estimates or significance testing. Single-run benchmark comparisons are insufficient for a claim of superiority.
minor comments (5)
  1. [Title and running header] The running title appears as 'AF-MA T' in the paper text; this typo should be fixed.
  2. [Case Study, Table 3] The column header of Table 3 appears garbled ('Text A TAE-LSTM IAN IA-GCN AF-MA T Labels'); the intended grouping of model columns is not clear.
  3. [MC2F Block, Eq. (15)] Equation (15) adds H to the output of a Linear layer, but the text does not state that the projection output dimension matches the hidden state dimension; this should be clarified.
  4. [Implementation Details] The text says 'We used 2 layers and 4 heads for AF-MAT+BERT (2 heads for AF-MAT)' but does not explicitly state the number of layers for the non-BERT AF-MAT; the L study suggests 2, so this should be stated directly.
  5. [Experiment, Datasets] The datasets are referred to inconsistently as 'Restaurant' and 'Rest14'; please unify the terminology throughout.

Circularity Check

2 steps flagged · score 6.0 of 10

Hyperparameter r/L are tuned on Restaurant14 before its score is reported, and the Eq. (4) aspect gate is token-independent and cancels in normalization; both central claims reduce by construction.

  1. fitted input called prediction [Effect of Hyperparameter r; Effect of Hyperparameter L; Implementation Details; Table 2]
    "In this section, we investigate the impact of r hyperparameter in the AF-MAT framework using the Restaurant dataset. ... we observe that AF-MAT achieves optimal performance when r = 6 ... the AF-MAT model achieves its highest performance on the Restaurant dataset when configured with two layers."

    The reported AF-MAT Rest14 accuracy (84.67, Table 2) is produced by a configuration whose hyperparameters r and L were selected by maximizing performance on that same Restaurant14 dataset (Figures 3 and 4). The paper describes no held-out validation split, so the Restaurant14 result is not an independent prediction of the chosen configuration; it is the value used to pick the configuration. Reporting it as evidence of SOTA status is therefore a fitted input renamed as a prediction, at least on the Restaurant14 benchmark where the tuning was performed.

  2. self definitional [Aspect-aware mLSTM (AA-mLSTM), Eqs. (4)-(11)]
    "The resulting vector ˜H fwd a serves as the global aspect representation and is broadcast across all time steps, which is used in two ways: (1) as input to a new aspect gate, and (2) as the key vector kt ... at = exp(Wa[ ˜Ha fwd ⊕ ˜Ha fwd ⊕ ˜Ha fwd ] + ba) ... αt,j = q⊤t kj√d ... eCt,j = Ct,j / Pt j=1 Ct,j + ϵ"

    The gate's RHS has no token index t; at is a per-sample constant. The text also states the broadcast vector is used as kt, so αt,j is independent of j. Eq. (7) subtracts the max over j, which removes any global constant, and Eq. (10) normalizes over j, which cancels any j-independent factor. Therefore eCt,j — and hence H_fwd_t in Eq. (11) — are identical with or without the aspect gate and aspect key. The claimed token-level 'selective emphasis' is not produced by these equations; the component that the paper credits for the Table 4 gains is a no-op by construction. The core AA-mLSTM contribution reduces to a constant that cancels, so the architecture's stated novelty cannot explain the reported results.

full rationale

The paper is primarily an empirical architecture paper; its own derivation chain is short and is mostly a forward-pass specification. Self-citations to Lawan et al. 2025 appear as baselines and related work, but they are empirical comparisons rather than load-bearing circular arguments, so they do not themselves raise the score. However, two reductions by construction undermine the central claims. First, r and L are selected by maximizing Restaurant14 accuracy, and the resulting Restaurant14 score is then reported as the headline non-BERT result without any described held-out split, so that number is partly fitted to the benchmark it claims to predict. Second, the defining equation of the aspect gate is token-independent, and the subsequent max-subtraction and normalization cancel any constant; the claimed aspect-aware gating is therefore a no-op as written, making the ablation conclusion and the SOTA attribution unrealizable from the stated equations. The missing 'Equation ??' for q_t, v_t, i_t, f_t is a specification gap that prevents reconstruction of the forward pass; it is a correctness and verifiability issue rather than circularity, but it compounds the problem. Because the two reductions concern the paper's central claims (best accuracy and the aspect-gate mechanism), the circularity score is 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the implementability of the modified mLSTM equations, transferability of tuned hyperparameters, and comparability of literature baselines. No code is provided, and Eq. (4) plus the unresolved 'Equation ??' leave the core mechanism under-specified.

free parameters (3)
  • partial flip length r = 6
    Tuned on Restaurant14 (Figure 3) and used for all final results; no held-out validation is described.
  • number of layers L = 2
    Tuned on Restaurant14 (Figure 4) and used for final model.
  • number of attention/fusion heads = 4 for AF-MAT+BERT, 2 for AF-MAT
    Set per variant in Implementation Details; not clear whether tuned per dataset.
assumptions (5)
  • standard math The mLSTM update equations from xLSTM remain valid when augmented with an aspect gate.
    The AA-mLSTM block (Eqs. 5-11) inherits the mLSTM formulation without re-derivation.
  • domain assumption Processing reversed token sequences preserves the semantic signals needed for ABSA.
    FlipMix (Eqs. 12-13) assumes that reverse-order processing adds useful context without distorting aspect-opinion associations.
  • ad hoc to paper The mean-pooled aspect vector is a sufficient conditioning signal for the aspect gate.
    Eq. (4) builds the gate from the pooled aspect vector only; the paper provides no evidence for sufficiency.
  • domain assumption Baseline numbers from the cited papers are directly comparable to the authors' runs.
    Table 2 mixes published results with the authors' implementations; no common protocol or code is provided.
  • ad hoc to paper Hyperparameters tuned on Restaurant14 transfer to Laptop14 and Twitter.
    No validation set is described; the same r=6 and L=2 are used for all datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AF-MAT: Aspect-aware Flip-and-Fuse xLSTM for Aspect-based Sentiment Analysis." pith.science (2026). https://pith.science/paper/ZOLDWA3P

@misc{pith2026250701213,
  author       = {Pith},
  title        = {Pith review of: AF-MAT: Aspect-aware Flip-and-Fuse xLSTM for Aspect-based Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZOLDWA3P}},
  note         = {Machine review of arXiv:2507.01213}
}
read the original abstract

Aspect-based Sentiment Analysis (ABSA) is a crucial NLP task that extracts fine-grained opinions and sentiments from text, such as product reviews and customer feedback. Existing methods often trade off efficiency for performance: traditional LSTM or RNN models struggle to capture long-range dependencies, transformer-based methods are computationally costly, and Mamba-based approaches rely on CUDA and weaken local dependency modeling. The recently proposed Extended Long Short-Term Memory (xLSTM) model offers a promising alternative by effectively capturing long-range dependencies through exponential gating and enhanced memory variants, sLSTM for modeling local dependencies, and mLSTM for scalable, parallelizable memory. However, xLSTM's application in ABSA remains unexplored. To address this, we introduce Aspect-aware Flip-and-Fuse xLSTM (AF-MAT), a framework that leverages xLSTM's strengths. AF-MAT features an Aspect-aware matrix LSTM (AA-mLSTM) mechanism that introduces a dedicated aspect gate, enabling the model to selectively emphasize tokens semantically relevant to the target aspect during memory updates. To model multi-scale context, we incorporate a FlipMix block that sequentially applies a partially flipped Conv1D (pf-Conv1D) to capture short-range dependencies in reverse order, followed by a fully flipped mLSTM (ff-mLSTM) to model long-range dependencies via full sequence reversal. Additionally, we propose MC2F, a lightweight Multihead Cross-Feature Fusion based on mLSTM gating, which dynamically fuses AA-mLSTM outputs (queries and keys) with FlipMix outputs (values) for adaptive representation integration. Experiments on three benchmark datasets demonstrate that AF-MAT outperforms state-of-the-art baselines, achieving higher accuracy in ABSA tasks.

Figures

Figures reproduced from arXiv: 2507.01213 by the authors.

Figure 1
Figure 1. An example sentence demonstrating the need for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The complete AF-MAT architecture for ABSA. The model integrates AA-mLSTM with a dedicated aspect gate, a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Effect of parameter r [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 39 canonical work pages

  1. [1]

    The battery life is great, but the screen is dim,

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

    work page 2025

  2. [2]

    Cambria, B

    E. Cambria, B. Schuller, Y. Xia, and C. Havasi, ``New avenues in opinion mining and sentiment analysis,'' Tech. Rep. [Online]. Available: http://converseon.com

    work page

  3. [3]

    D. Tang, B. Qin, and T. Liu, ``Aspect level sentiment classification with deep memory network,'' in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 2016, pp. 214--224

    work page 2016

  4. [4]

    Y. Wang, M. Huang, L. Zhao, and X. Zhu, ``Attention-based lstm for aspect-level sentiment classification,'' in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 2016, pp. 606--615

    work page 2016

  5. [5]

    D. Ma, S. Li, X. Zhang, and H. Wang, ``Interactive attention networks for aspect-level sentiment classification,'' in IJCAI'17: Proceedings of the 26th International Joint Conference on Artificial Intelligence, 9 2017. [Online]. Available: http://arxiv.org/abs/1709.00893

    arXiv 2017

  6. [6]

    C. Peng, S. Zhongqian, B. Lidong, and W. Yang, ``Recurrent attention network on memory for aspect sentiment analysis,'' in Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 2017, pp. 452--461

    work page 2017

  7. [7]

    F. Fan, Y. Feng, and D. Zhao, ``Multi-grained attention network for aspect-level sentiment classification,'' in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 2018, pp. 3433--3442

    work page 2018

  8. [8]

    C. Yang, H. Zhang, B. Jiang, and K. Li, ``Aspect-based sentiment analysis with alternating coattention networks,'' Information Processing and Management, vol. 56, pp. 463--478, 5 2019

    work page 2019

Show all 49 references
  1. [9]

    Liu and B

    N. Liu and B. Shen, ``Aspect-based sentiment analysis with gated alternate neural network,'' vol. 188, p. 105010, 2020. [Online]. Available: https://doi.org/10.1016/j.knosys

    2020 doi

  2. [10]

    R. K. Yadav, L. Jiao, M. Goodwin, and O. C. Granmo, ``Positionless aspect based sentiment analysis using attention mechanism[formula presented],'' Knowledge-Based Systems, vol. 226, 8 2021

    2021

  3. [11]

    X. Wang, M. Tang, T. Yang, and Z. Wang, ``A novel network with multiple attention mechanisms for aspect-level sentiment analysis,'' Knowledge-Based Systems, vol. 227, 9 2021

    2021

  4. [12]

    K. Sun, R. Zhang, S. Mensah, Y. Mao, and X. Liu, ``Aspect-level sentiment analysis via convolution over dependency tree,'' in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language P...

    2019

  5. [13]

    Zhang, Q

    C. Zhang, Q. Li, and D. Song, ``Aspect-based sentiment classification with aspect-specific graph convolutional networks,'' in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language P...

    2019

  6. [14]

    Liang, H

    B. Liang, H. Su, L. Gui, E. Cambria, and R. Xu, ``Aspect-based sentiment analysis via affective knowledge enhanced graph convolutional networks,'' Knowledge-Based Systems, vol. 235, 1 2022

    2022

  7. [15]

    H. Wu, C. Huang, and S. Deng, ``Improving aspect-based sentiment analysis with knowledge-aware dependency graph network,'' Information Fusion, vol. 92, pp. 289--299, 4 2023

    2023

  8. [16]

    T. Gu, H. Zhao, Z. He, M. Li, and D. Ying, ``Integrating external knowledge into aspect-based sentiment analysis using graph neural network,'' Knowledge-Based Systems, vol. 259, 1 2023

    2023

  9. [17]

    H. Liu, Y. Wu, Q. Li, W. Lu, X. Li, J. Wei, X. Liu, and J. Feng, ``Enhancing aspect-based sentiment analysis using a dual-gated graph convolutional network via contextual affective knowledge,'' Neurocomputing, vol. 553, 10 2023

    2023

  10. [18]

    P. Li, P. Li, and X. Xiao, ``Aspect-pair supervised contrastive learning for aspect-based sentiment analysis,'' Knowledge-Based Systems, vol. 274, 8 2023

    2023

  11. [19]

    X. Song, G. Ling, W. Tu, and Y. Chen, ``Knowledge-guided heterogeneous graph convolutional network for aspect-based sentiment analysis,'' Electronics (Switzerland), vol. 13, 2 2024

    2024

  12. [20]

    Y. Tay, L. A. Tuan, and S. C. Hui, ``Learning to attend via word-aspect associative fusion for aspect-based sentiment analysis,'' in Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Con...

    2018

  13. [21]

    Hazarika, S

    D. Hazarika, S. Poria, P. Vij, G. Krishnamurthy, E. Cambria, and R. Zimmermann, ``Modeling inter-aspect dependencies for aspect-based sentiment analysis,'' in Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Hu...

    2018

  14. [22]

    M. H. Phan and P. Ogunbona, ``Modelling context and syntactical features for aspect-based sentiment analysis,'' Tech. Rep. [Online]. Available: https://universaldependencies.org/u/pos/

  15. [23]

    C. Chen, Z. Teng, Z. Wang, and Y. Zhang, ``Discrete opinion tree induction for aspect-based sentiment analysis,'' Tech. Rep. [Online]. Available: https://github.com/CCSoleil/dotGCN

  16. [24]

    Gu and T

    A. Gu and T. Dao, ``Mamba: Linear-time sequence modeling with selective state spaces,'' 12 2023. [Online]. Available: http://arxiv.org/abs/2312.00752

    2023 arXiv

  17. [25]

    Lawan, J

    A. Lawan, J. Pu, H. Yunusa, A. Umar, and M. Lawan, ``Enhancing long-range dependency with state space model and kolmogorov-arnold networks for aspect-based sentiment analysis,'' in Proceedings of the 31st International Conference on Computational Linguistics, O. Rambow, L. Wan...

    2025

  18. [26]

    M. Beck, K. Pöppel, M. Spanring, A. Auer, O. Prudnikova, M. Kopp, G. Klambauer, J. Brandstetter, and S. Hochreiter, ``xlstm: Extended long short-term memory,'' 12 2024. [Online]. Available: http://arxiv.org/abs/2405.04517

    2024 arXiv

  19. [27]

    C. Zhu, B. Yi, and L. Luo, ``Aspect-based sentiment analysis via bidirectional variant spiking neural p systems,'' Expert Systems with Applications, vol. 259, p. 125295, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0957417424021626

    2025

  20. [28]

    Ouyang, C

    J. Ouyang, C. Xuan, B. Wang, and Z. Yang, ``Aspect-based sentiment classification with aspect-specific hypergraph attention networks,'' Expert Systems with Applications, vol. 248, p. 123412, 8 2024

    2024

  21. [29]

    Wu and G

    Y. Wu and G. Deng, ``Aspect-level sentiment analysis based on graph convolutional networks and interactive aggregate attention,'' Computer Speech and Language, vol. 95, 1 2026

    2026

  22. [30]

    S. Feng, B. Wang, Z. Yang, and J. Ouyang, ``Aspect-based sentiment analysis with attention-assisted graph and variational sentence representation,'' Knowledge-Based Systems, vol. 258, 12 2022

    2022

  23. [31]

    B. Yu, C. Cao, and Y. Yang, ``Dynamic position weighting aspect-focused graph convolutional network for aspect-based sentiment analysis,'' Journal of Supercomputing, vol. 81, 1 2025

    2025

  24. [32]

    F. Luo, A. Li, B. Jiang, S. Khan, K. Wu, and L. Wang, ``Activitymamba: A cnn-mamba hybrid neural network for efficient human activity recognition,'' IEEE Transactions on Mobile Computing, 2025

    2025

  25. [33]

    A. Gu, T. Dao, S. Ermon, A. Rudra, and C. Ré, ``Hippo: Recurrent memory with optimal polynomial projections,'' in 34th Conference on Neural Information Processing Systems (NeurIPS 2020), 2020. [Online]. Available: https://github.com/HazyResearch/hippo-code

    2020

  26. [34]

    A. Gu, K. Goel, and C. Ré, ``Efficiently modeling long sequences with structured state spaces,'' in ICLR 2022, 10 2021. [Online]. Available: http://arxiv.org/abs/2111.00396

    2022 arXiv

  27. [35]

    Y. Kong, Z. Wang, Y. Nie, T. Zhou, S. Zohren, Y. Liang, P. Sun, and Q. Wen, ``Unlocking the power of lstm for long term time series forecasting,'' Tech. Rep., 2025. [Online]. Available: https://github.com/Eleanorkong/P-sLSTM

    2025

  28. [36]

    Z. Wu, X. Ma, R. Lian, Z. Lin, and W. Zhang, ``Cdxformer: Boosting remote sensing change detection with extended long short-term memory,'' 11 2024. [Online]. Available: http://arxiv.org/abs/2411.07863

    2024 arXiv

  29. [37]

    X. He, W. Ni, Z. Zhang, H. Luo, and L. Wan, ``Msgcn-xlstm: Efficient wind power forecasting approach combining multi-scale graph convolutional network and extended lstm,'' IEEE Sensors Journal, 2025

    2025

  30. [38]

    Alkin, M

    B. Alkin, M. Beck, K. Pöppel, S. Hochreiter, and J. Brandstetter, ``Vision-lstm: xlstm as generic vision backbone,'' 2025. [Online]. Available: https://arxiv.org/abs/2406.04303

    2025 arXiv

  31. [39]

    N. L. Kühne, J. Østergaard, J. Jensen, and Z.-H. Tan, ``xlstm-senet: xlstm for single-channel speech enhancement,'' 5 2025. [Online]. Available: http://arxiv.org/abs/2501.06146

    2025 arXiv

  32. [40]

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, ``Vision mamba: Efficient visual representation learning with bidirectional state space model,'' 1 2024. [Online]. Available: http://arxiv.org/abs/2401.09417

    2024 arXiv

  33. [41]

    J. Zhu, X. Chen, K. He, Y. LeCun, and Z. Liu, ``Transformers without normalization,'' in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), June 2025, pp. 14\,901--14\,911

    2025

  34. [42]

    Y.-H. H. Tsai, S. Bai, P. P. Liang, J. Z. Kolter, L.-P. Morency, and R. Salakhutdinov, ``Multimodal transformer for unaligned multimodal language sequences,'' in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, A. Korhonen, D. Traum, and...

    2019

  35. [43]

    X. He, K. Cao, J. Zhang, K. Yan, Y. Wang, R. Li, C. Xie, D. Hong, and M. Zhou, ``Pan-mamba: Effective pan-sharpening with state space model,'' Information Fusion, vol. 115, p. 102779, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1566253524005578

    2025

  36. [44]

    Pontiki, H

    M. Pontiki, H. Papageorgiou, D. Galanis, I. Androutsopoulos, J. Pavlopoulos, and S. Manandhar, ``Semeval-2014 task 4: Aspect based sentiment analysis,'' in Proceedings of the 8th International Workshop on Semantic Evaluation, 2014, pp. 27--35. [Online]. Available: http://alt.qcri

    2014

  37. [45]

    L. Dong, F. Wei, C. Tan, D. Tang, M. Zhou, and K. Xu, ``Adaptive recursive neural network for target-dependent twitter sentiment classification,'' in Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics. 1em plus 0.5em minus 0.4em Association...

    2014

  38. [46]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, ``Bert: Pre-training of deep bidirectional transformers for language understanding,'' in Proceedings of NAACL-HLT 2019, 10 2018. [Online]. Available: http://arxiv.org/abs/1810.04805

    2019 arXiv

  39. [47]

    Y. Song, J. Wang, T. Jiang, Z. Liu, and Y. Rao, ``Attentional encoder network for targeted sentiment classification,'' 2 2019. [Online]. Available: http://arxiv.org/abs/1902.09314 http://dx.doi.org/10.1007/978-3-030-30490-4_9

    2019 arXiv

  40. [48]

    Zhang, Q

    C. Zhang, Q. Li, and D. Song, ``Aspect-based sentiment classification with aspect-specific graph convolutional networks,'' pp. 4568--4578, 2019. [Online]. Available: https://spacy.io/

    2019

  41. [49]

    H. Tang, D. Ji, C. Li, and Q. Zhou, ``Dependency graph enhanced dual-transformer structure for aspect-based sentiment classification,'' in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistic, 2020, pp. 6578--6588

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.