Pith. sign in

REVIEW 5 major objections 7 minor 48 references

Quantum Graph Transformer for NLP Sentiment Classification

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing the learned query/key projections of a graph transformer with parameterized quantum circuits yields higher accuracy than an equivalent classical graph transformer on five sentiment benchmarks, while using…

desk verdict Plausible new QGT architecture, but headline gains over an underspecified classical baseline; needs baseline specs and code before claims are believable. read the letter →

arxiv 2506.07937 v1 pith:MRMOOQZR submitted 2025-06-09 cs.CL quant-ph

Shamminuj Aktar , Andreas Bärtschi , Abdel-Hameed A. Badawy , Stephan Eidenbenz This is my paper
classification cs.CLquant-ph MSC 68T5081P6868T07
keywords quantummachinelearninggraphtransformerself-attentionsentimentclassificationparameterizedcircuitsnaturallanguageprocessing
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

This paper introduces the Quantum Graph Transformer (QGT), a hybrid quantum-classical model for sentiment classification. QGT represents each sentence as a fully connected graph of tokens and uses parameterized quantum circuits (PQCs) to compute the query and key vectors for self-attention, which then guide message passing. The authors claim that QGT matches or exceeds existing quantum NLP models and beats an equivalent classical graph transformer by 5.42% average accuracy on real-world datasets and 4.76% on synthetic ones. They also report that QGT reaches 82% accuracy using only 30% of the Yelp training samples, while the classical model needs 60%. If these results hold, QGT offers a small-scale demonstration that quantum-generated attention features can be competitive with classical ones at a fraction of the parameter count.

What carries the argument

The load-bearing component is the QTransformerConv layer, a message-passing step in which attention scores come from quantum-computed query and key vectors rather than learned linear projections. Each token embedding is loaded into a 4-qubit state using grouped Ry rotations plus a ring of CNOT entanglements; the query circuit U_Q($\theta$) and key circuit U_K(phi) apply alternating Rx, Ry, Rz rotations, a CNOT ring, a final Ry layer, and a Quantum Fourier Transform. Query and key vectors are the expectation values of Pauli-X measurements on the processed states, and the scaled dot product of these vectors defines the attention coefficients used to aggregate neighbor messages. This design replaces the classical Q and K weight matrices with 32 PQC parameters, which is what the paper credits for both parameter reduction and the reported accuracy and sample-efficiency gains.

What would settle it

Re-run the classical graph transformer with the same token embedding dimension (d=16), the same number of layers, the same Adam optimizer (learning rate 0.01), batch size 32, early stopping, and the same 70/10/20 data splits as QGT, and report its test accuracy and parameter count. If the classical model closes the 5.42% gap on Yelp, IMDB, and Amazon to within one standard deviation, the claimed quantum advantage is not supported.

Watch

Extended reading notes

Core claim

The central claim is that a graph transformer whose self-attention queries and keys are produced by parameterized quantum circuits—using just 32 trainable parameters per layer for both query and key generation—can outperform a classical graph transformer on sentiment classification. QGT encodes each token's 16-dimensional GloVe-derived embedding into a 4-qubit state, applies separate PQC layers to extract query and key vectors via Pauli-X expectation values, computes scaled dot-product attention, and uses the resulting scores for message passing. The authors report average accuracy improvements of 5.42% on Yelp, IMDB, and Amazon and 4.76% on the MC and RP synthetic datasets, with a nearly 50% reduction in required labeled samples on Yelp. The paper interprets this as evidence that quantum-enhanced self-attention can capture contextual relationships in language with significantly fewer trainable parameters than classical attention mechanisms.

Load-bearing premise

The 5.42% and 4.76% accuracy improvements rest on the assumption that the 'equivalent classical graph transformer' baseline is genuinely equivalent and fairly tuned, but the paper gives no architecture details, parameter counts, or hyperparameter settings for that baseline, so the margin could partly reflect an under-tuned classical comparison rather than a genuine quantum advantage.

Editorial extensions

If this is right

  • QGT uses only 32 trainable parameters for its quantum self-attention layers, compared with the high-dimensional projection matrices of classical attention, so the architecture directly reduces the parameter footprint of the attention mechanism.
  • On the Yelp dataset, QGT reaches about 82% test accuracy using 30% of the training samples, whereas the classical graph transformer requires 60% of samples for similar accuracy, indicating improved sample efficiency at small scale.
  • QGT achieves 100% accuracy on both synthetic MC and RP datasets, matching grammar-aware DisCoCat baselines, and consistently outperforms the attention-based QNLP models QSAN, QMSAN, and QSANN on real-world datasets.
  • The graph design is modular: the paper states that k-nearest-neighbor graphs, multi-head attention, and multiple PQC layers can be added for larger datasets, with qubit count growing linearly with the number of heads.
  • If the reported margins are stable, QGT provides a concrete hybrid architecture that can be trained end-to-end on a classical simulator using PennyLane, with backpropagation flowing through both quantum and classical parameters.

Reading between the lines

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

  • The sample-efficiency result would be more convincing if the classical baseline were matched not only in task and data but also in attention parameter budget (e.g., a classical bottleneck attention with 32 parameters); a re-run with such a control would isolate the quantum contribution from the effect of fewer parameters.
  • The experiments run entirely on PennyLane's classical simulator, so the reported accuracy is noise-free and ignores shot noise and gate errors; a natural extension is to execute QGT on real 4-qubit hardware and measure the accuracy degradation, which would test whether the advantage survives noise.
  • Because QGT uses fully connected token graphs, its asymptotic attention cost is O(N^2) exactly like classical transformers; the claimed advantage is therefore constant-factor (parameter count and sample efficiency), not a scaling improvement, so the fair scalability comparison is against classical models with matched parameter budgets rather than against full-scale pretrained transformers.
  • The attention heatmap in the paper suggests the model gravitates toward sentiment-bearing words, but the paper does not quantify whether attention weights correlate with human judgments; a testable follow-up is to compare QGT attention maps with LIME or SHAP attributions on the same sentences.
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 / 7 minor

Summary. The paper proposes the Quantum Graph Transformer (QGT), a hybrid quantum-classical architecture that replaces the query/key computation in a graph transformer's self-attention with parameterized quantum circuits (PQCs) and integrates this into a message-passing framework for sentiment classification. The model is evaluated on five small datasets (Yelp, IMDB, Amazon, MC, RP) and compared with several QNLP baselines and a 'classical graph transformer' baseline. The central claims are that QGT achieves higher or comparable accuracy than existing QNLP models, improves average accuracy by 5.42% on real-world and 4.76% on synthetic datasets over the classical baseline, and requires nearly 50% fewer labeled samples on Yelp. The paper also reports parameter reduction from using PQCs for attention.

Significance. If the reported accuracy and sample-efficiency advantages are substantiated, QGT would be a competitive hybrid architecture for small-scale sentiment classification and a useful contribution to the QNLP literature. The design is clearly described: sentence-to-graph construction, quantum encoding, PQC-based query/key generation, and message passing are all specified in enough detail to be reproduced, and the training/validation curves indicate stable convergence on the three real-world datasets. However, the quantitative headline claims currently rest on an underspecified classical baseline and on statistical reporting that is not sufficiently rigorous, so the significance is conditional on these issues being resolved.

major comments (5)
  1. [Section 4.3, Figure 6] The 'classical graph transformer' baseline (also labeled 'Classical GNN' in Figure 7) is not specified in terms of architecture, parameter count, hyperparameters, optimization budget, or graph construction. The abstract's claims of 5.42% and 4.76% average accuracy improvements and the sample-efficiency claim in Section 4.4 all derive directly from this comparison. As written, the comparison does not rule out the possibility that the classical baseline is undertuned (e.g., fixed learning rate, insufficient width, or a different training schedule), which would make the observed gaps an artifact rather than evidence of quantum benefit. Please provide a full experimental specification of the baseline, including its parameter count and hyperparameter search, or release code for both models.
  2. [Section 4.1] The QGT model is trained with an additional 'reinforcement learning-based regularization' in which the reward is the negative training loss and is used to directly update the PQC parameters. No analogous mechanism is described for the classical graph transformer baseline. This introduces a training-signal confound: the accuracy gap between QGT and the classical baseline could be partly due to this extra RL update rather than to the quantum attention mechanism. Please clarify how the RL update is combined with the Adam/cross-entropy gradient, and either apply an equivalent auxiliary update to the classical baseline or remove the RL term from the comparison.
  3. [Section 3.3, Eqs. (9)-(11)] The query and key vectors Q_i and K_i defined in Eqs. (9)-(10) are n-dimensional (n=4 qubits, one expectation value per qubit), but the attention scaling in Eq. (11) uses sqrt(d) with d=16. The scaling factor should be consistent with the actual dimension of the query/key vectors, i.e., sqrt(4), unless the PQC is designed to produce d-dimensional outputs. Please correct this discrepancy or clarify the intended dimension.
  4. [Section 4.4, Figure 7] The sample-efficiency claim is based on a single curve for one dataset (Yelp) with no error bars, and the claim of 'nearly 50% fewer labeled samples' depends on a specific threshold (82% accuracy at 30% vs. 60% of training data) that is not justified. Please report multiple random initializations/seeds, show variance, and, ideally, include at least one additional dataset to support a general sample-efficiency claim.
  5. [Section 4.3, Figure 6] The paper reports standard deviations (e.g., '93.0 ± 1.5') but never states the number of independent runs used to compute them. Without this information, the error bars are uninterpretable, and the claim that QGT 'consistently achieves higher' accuracy cannot be statistically assessed. Please specify the number of seeds and the method used to compute the standard deviations.
minor comments (7)
  1. [Section 2.4, Eq. (1)] Equation (1) states that W_Q, W_K, W_V are in R^d, but as projection matrices they should be in R^{d×d} (or R^{d'×d}); please correct the dimensionality.
  2. [Section 3, first paragraph] There are several typos: 'hybird' should be 'hybrid', 'intergrates' should be 'integrates', 'nerual' should be 'neural'. Please proofread the manuscript.
  3. [Section 3.3] The text says 'we apply a Pauli-X measurement on each qubit' but does not specify the number of measurement shots used to estimate the expectation values in Eqs. (9) and (10). Please state the shot count or clarify whether exact expectation values are simulated.
  4. [Figure 5] The attention heatmap shows several rows that are entirely zero; please clarify whether these correspond to padding tokens or to a systematic property of the attention mechanism.
  5. [Section 4.3 and Figure 7] The manuscript alternates between 'classical graph transformer' and 'Classical GNN' for the same baseline; please use one consistent name throughout.
  6. [References] Reference [8] (Church, 2017) is cited for Word2Vec and also appears as 'DiscoCat [8, 21]' in Figure 6; the correct citation for DisCoCat is [9] (Coecke et al., 2010). Please fix the citation mapping.
  7. [Section 4.1] Please report the total number of trainable parameters for the QGT model (quantum plus classical) and, for comparison, the parameter count of the classical graph transformer baseline; the current claim of parameter reduction in Section 4.4 is not quantified against the actual baseline.

Circularity Check

0 steps flagged · score 0.0 of 10

No equation-level or self-citation circularity; central claims are empirical evaluations against external benchmarks.

full rationale

The paper's derivation chain is not circular. QGT's quantum self-attention is defined through explicit circuits and equations (Eqs. 5-11), where query and key vectors are expectation values of Pauli observables from parameterized circuits. The attention formula (Eq. 11) is structurally the classical softmax scaled dot-product, but its inputs are computed by the PQC rather than by classical projection matrices, so the architecture is a genuine hybrid construction rather than a renamed version of the baseline. The accuracy and sample-efficiency claims come from training on five external benchmark datasets (Yelp, IMDB, Amazon, MC, RP) and comparing against reported values from other groups' QNLP models and an in-house classical graph transformer. Those results are empirical, not derived from assumptions that include the conclusions. The parameter-reduction claim (32 trainable parameters for U_Q and U_K with 4 qubits) is a design property of the chosen circuit, not a fitted input that is then relabeled as a prediction. The underspecification of the 'equivalent classical graph transformer' baseline in Section 4.3/Figure 6 is a legitimate concern about verifiability and whether the baseline was fairly tuned, but the paper does not state that the baseline is derived from QGT or that QGT's superiority is forced by construction. There is no load-bearing self-citation chain: the cited QNLP baselines (QSAN, QMSAN, QSANN, LEXIQL, DisCoCat) are external works, and no uniqueness theorem or prior result by the present authors is invoked to rule out alternatives. The RL-based regularization described in Section 4.1 is an additional training signal for QGT only, and while it may partly explain the gains over the classical baseline, that is a training-protocol difference, not circular reasoning. Accordingly, no specific circular step can be quoted with an exhibiting reduction, and the honest finding is 'no significant circularity' (score 0).

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The model relies on trained PQC parameters, trained classical projections, and hand-chosen hyperparameters, plus domain assumptions about GloVe embeddings, fully connected graph structure, and ideal circuit simulation. No new physical entities are posited.

free parameters (6)
  • theta (query PQC parameters) = trained on each dataset
    16 parameters in U_Q(theta) learned during training, described in Section 3.3.
  • phi (key PQC parameters) = trained on each dataset
    16 parameters in U_K(phi) learned during training, described in Section 3.3.
  • embedding projection weights = trained on each dataset
    Linear layer maps GloVe embeddings from dimension 50 to dimension 16, Section 3.2.
  • value projection weights = trained on each dataset
    Projects node features to value vectors used in message passing, Section 3.4.
  • classification head weights and bias = trained on each dataset
    Fully connected layer from pooled representation to class logits, Section 3.5.
  • model hyperparameters (qubits=4, d=16, lr=0.01, batch=32, epochs=25) = chosen by hand
    Stated in Section 4.1 without a sensitivity study or search procedure.
assumptions (3)
  • domain assumption GloVe embeddings provide useful semantic initialization for tokens.
    Invoked in Section 3.2; no ablation or sensitivity analysis is given.
  • domain assumption A fully connected token graph captures the contextual dependencies needed for sentiment classification.
    Used in Section 3.1; the alternative k-NN graph is mentioned but not evaluated.
  • domain assumption PennyLane Lightning accurately simulates ideal noiseless quantum circuits.
    All experiments use simulation on a Mac with PennyLane Lightning, Section 4.1; noise effects are deferred to future work in Section 5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum Graph Transformer for NLP Sentiment Classification." pith.science (2026). https://pith.science/paper/MRMOOQZR

@misc{pith2026250607937,
  author       = {Pith},
  title        = {Pith review of: Quantum Graph Transformer for NLP Sentiment Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRMOOQZR}},
  note         = {Machine review of arXiv:2506.07937}
}
read the original abstract

Quantum machine learning is a promising direction for building more efficient and expressive models, particularly in domains where understanding complex, structured data is critical. We present the Quantum Graph Transformer (QGT), a hybrid graph-based architecture that integrates a quantum self-attention mechanism into the message-passing framework for structured language modeling. The attention mechanism is implemented using parameterized quantum circuits (PQCs), which enable the model to capture rich contextual relationships while significantly reducing the number of trainable parameters compared to classical attention mechanisms. We evaluate QGT on five sentiment classification benchmarks. Experimental results show that QGT consistently achieves higher or comparable accuracy than existing quantum natural language processing (QNLP) models, including both attention-based and non-attention-based approaches. When compared with an equivalent classical graph transformer, QGT yields an average accuracy improvement of 5.42% on real-world datasets and 4.76% on synthetic datasets. Additionally, QGT demonstrates improved sample efficiency, requiring nearly 50% fewer labeled samples to reach comparable performance on the Yelp dataset. These results highlight the potential of graph-based QNLP techniques for advancing efficient and scalable language understanding.

Figures

Figures reproduced from arXiv: 2506.07937 by the authors.

Figure 1
Figure 1. Overview of the Quantum Graph Transformer (QGT) architecture. Each input sentence S is tokenized and represented as a fully connected graph 𝐺. Tokens are initialized using GloVe embeddings (𝑑0 = 50) and projected to x𝑖 ∈ R 𝑑 with 𝑑 = 16, then processed through stacked QTransformerConv layers and aggregated via global mean pooling. In each QTransformerConv layer, node features x𝑖 are encoded into quantum states |𝜓𝑖⟩ … view at source ↗
Figure 2
Figure 2. (Left) Quantum embedding circuit 𝑈encoding (x𝑖) maps a token embedding vector x𝑖 ∈ R 𝑑=16 to a 4-qubit quantum state |𝜓𝑖⟩ using grouped 𝑅𝑦 rotations, where each rotation angle is given by Í4𝑘+3 𝑗=4𝑘 x 𝑗 𝑖 for qubits 𝑘 = 0, 1, 2, 3. The circuit is then followed by a ring-style CNOT entanglement layer. (Right) Parameterized quantum circuit 𝑈𝑄 (𝜽) used for query vector generation consists of 𝑅𝑥 , 𝑅𝑦 and 𝑅𝑧 layers per q… view at source ↗
Figure 3
Figure 3. Training loss curves for the Yelp (left), IMDB (middle), and Amazon (right) datasets in the sentiment classification task using the proposed QGT model. All three datasets show consistent loss reduction /learning over training epochs. 0 5 10 15 20 25 Epochs 0 20 40 60 80 100 Accuracy (%) Yelp Dataset Validation Accuracy Training Accuracy 0 5 10 15 20 25 Epochs 0 20 40 60 80 100 IMDB Dataset Validation Accuracy Traini… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Training and validation accuracy curves for the Yelp (left), IMDB (middle), and Amazon (right) datasets in the sentiment classification task using the proposed QGT model. The consistent gap between training and validation curves indicates stable convergence and strong …
Figure 5
Figure 5. Figure 5: Attention score heatmap of a testing sample from the Amazon dataset. Lighter colors indicate higher attention given to the corresponding token from other tokens. applies the softmax function internally to compute the prob￾ability distribution over classes. The loss is …
Figure 6
Figure 6. Figure 6: (Left) Test accuracy comparison between the proposed QGT model and existing attention-based models, including a classical graph transformer model. (Right) Test accuracy comparison between the QGT model and other baseline models. The QGT model consistently outperforms a…
Figure 7
Figure 7. Figure 7: Demonstrating the effect of varying sample sizes from the Yelp dataset on model learning performance. The plot shows accuracy on a fixed testing set for both the QGT model and a classical graph transformer. The QGT model demonstrates better learning efficiency with few…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 17 canonical work pages

  1. [1]

    Aaranya Alexander and Dominic Widdows. 2022. Quantum Text Encoding for Classification Tasks. In 2022 IEEE/ACM 7th Symposium on Edge Computing (SEC) . IEEE, 355–361. doi:10.1109/sec54971.2022. 00052 arXiv:2301.03715

    arXiv 2022

  2. [2]

    Ville Bergholm, Josh Izaac, et al. 2022. Pennylane: Automatic differen- tiation of hybrid quantum-classical computations. arXiv:1811.04968 arXiv preprint

    arXiv 2022

  3. [3]

    Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Wiebe, and Seth Lloyd. 2017. Quantum machine learn- ing. Nature 549 (Sept. 2017), 195–202. doi: 10.1038/nature23474 arXiv:1611.09347

    arXiv 2017

  4. [4]

    Brown, Vincent J

    Peter F. Brown, Vincent J. Della Pietra, Peter V. Desouza, Jennifer C. Lai, and Robert L. Mercer. 1992. Class-Based n-gram Models of Natural Language. Computational Linguistics 18, 4 (1992), 467–480. https: //aclanthology.org/J92-4003/

    work page 1992

  5. [5]

    Marco Cerezo, Guillaume Verdon, Hsin-Yuan Huang, Lukasz Cincio, and Patrick J. Coles. 2022. Challenges and opportunities in quantum machine learning. Nature Computational Science 2 (Sept. 2022), 567–

    work page 2022

  6. [6]

    Fu Chen, Qinglin Zhao, Li Feng, Chuangtao Chen, Yangbin Lin, and Jianhong Lin. 2025. Quantum mixed-state self-attention network. Neural Networks 185 (May 2025), 107123. doi: 10.1016/j.neunet.2025. 107123 arXiv:2403.02871

    work page Pith review arXiv 2025

  7. [7]

    Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio

    work page

  8. [8]

    Kenneth Ward Church. 2017. Word2Vec.Natural Language Engineering 23, 1 (Jan. 2017), 155–162. doi:10.1017/S1351324916000334

    doi:10.1017/s1351324916000334 2017

Show all 48 references
  1. [9]

    Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen Clark. 2010. Math- ematical Foundations for a Compositional Distributional Model of Meaning. Linguistic Analysis 36, 1–4 (March 2010), 345–384. arXiv:1003.4394 Festschrift for Joachim Lambek

    2010 arXiv

  2. [10]

    PyTorch Contributors. 2025. CrossEntropyLoss. https://pytorch. org/docs/stable/generated/torch.nn.CrossEntropyLoss.html Accessed: 2025-05-07

    2025

  3. [11]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

  4. [12]

    Xianzhong Ding, Le Chen, Murali Emani, Chunhua Liao, Pei-Hung Lin, Tristan Vanderbruggen, Zhen Xie, Alberto Cerpa, and Wan Du. 2023. HPC-GPT: Integrating Large Language Model for High-Performance Computing. In SC-W’23 Workshops of The International Conference on High Performan...

    2023

  5. [13]

    Luciano Floridi and Massimo Chiriatti. 2020. GPT-3: Its Nature, Scope, Limits, and Consequences. Minds and Machines 30 (Nov. 2020), 681–

    2020

  6. [14]

    Zellig S. Harris. 1954. Distributional Structure. Word 10, 2–3 (1954), 146–162. doi:10.1080/00437956.1954.11659520

    1954

  7. [15]

    Córcoles, Kristan Temme, Aram W

    Vojtěch Havlíček, Antonio D. Córcoles, Kristan Temme, Aram W. Harrow, Abhinav Kandala, Jerry M. Chow, and Jay M. Gambetta

  8. [16]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation 9, 8 (Nov. 1997), 1735–1780. doi:10.1162/ neco.1997.9.8.1735

    1997

  9. [17]

    Dimitrios Kotzias, Misha Denil, Nando De Freitas, and Padhraic Smyth

  10. [18]

    Guangxi Li, Xuanqiang Zhao, and Xin Wang. 2024. Quantum self- attention neural networks for text classification. Science China In- formation Sciences 67 (March 2024), 142501. doi:10.1007/s11432-023- 3879-7 arXiv:2205.05625

    2024 arXiv

  11. [19]

    Bang Liu and Lingfei Wu. 2022. Graph Neural Networks: Foundations, Frontiers, and Applications. Springer, Chapter Graph Neural Networks in Natural Language Processing, 463–481. doi: 10.1007/978-981-16- 6054-2_21

    2022 doi

  12. [20]

    Nature 567 (March 2019), 209–212

    Supervised learning with quantum-enhanced feature spaces. Nature 567 (March 2019), 209–212. doi: 10.1038/s41586-019-0980-2 arXiv:1804.11326

    2019 arXiv

  13. [21]

    Robin Lorenz, Anna Pearson, Konstantinos Meichanetzidis, Dim- itri Kartsaklis, and Bob Coecke. 2023. QNLP in Practice: Run- ning Compositional Models of Meaning on a Quantum Computer. Journal of Artificial Intelligence Research 76 (April 2023), 1305–1342. doi:10.1613/jair.1.14...

    2023 arXiv

  14. [22]

    Kelleher, and Andy Way

    Yasmin Moslem, Rejwanul Haque, John D. Kelleher, and Andy Way

  15. [23]

    Lakshmipathi Narayan. 2018. IMDb Dataset of 50K Movie Re- views. https://www.kaggle.com/datasets/lakshmi25npathi/imdb- dataset-of-50k-movie-reviews Accessed: 2025-05-14. Shamminuj Aktar, Andreas Bärtschi, Abdel-Hameed A. Badawy, and Stephan Eidenbenz

    2018

  16. [24]

    OpenAI, Josh Achiam, et al . 2024. GPT-4 Technical Report. arXiv preprint (March 2024), 1–100. arXiv:2303.08774

    2024 arXiv

  17. [25]

    Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global Vectors for Word Representation. In2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) . Associa- tion for Computational Linguistics, 1532–1543. doi: 10.3115/v1/d14- 1162

    2014 doi

  18. [26]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoy- anov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Ap- proach. arXiv preprint (July 2019), 1–13. arXiv:1907.11692

    2019 arXiv

  19. [27]

    Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. SemEval-2017 Task 4: Sentiment Analysis in Twitter. In 11th International Workshop on Semantic Evaluation (SemEval-2017). Association for Computational Linguistics, 502–518. doi:10.18653/v1/s17-2088 arXiv:1912.00741

    2017 arXiv

  20. [28]

    Russell and Peter Norvig

    Stuart J. Russell and Peter Norvig. 2016.Artificial Intelligence: A Modern Approach (4th ed.). Pearson. http://aima.cs.berkeley.edu/

    2016

  21. [29]

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. 2009. The Graph Neural Network Model. IEEE Transactions on Neural Networks 20, 1 (Jan. 2009), 61–80. doi:10. 1109/TNN.2008.2005605

    2009

  22. [30]

    Maria Schuld and Nathan Killoran. 2019. Quantum Machine Learning in Feature Hilbert Spaces. Physical Review Letters 122, 4 (Feb. 2019), 040504. doi:10.1103/physrevlett.122.040504 arXiv:1803.07128

    2019 arXiv

  23. [31]

    Daniel Silver, Aditya Ranjan, Rakesh Achutha, Tirthak Patel, and Devesh Tiwari. 2024. LEXIQL: Quantum Natural Language Processing on NISQ-era Machines. In SC’24: International Conference for High Performance Computing, Networking, Storage and Analysis . IEEE, IEEE, 1–15. doi:1...

    2024 arXiv

  24. [32]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. In 2013 Conference on Empirical Methods in Natural Language Process- ing (EMNLP...

    2013 doi

  25. [33]

    Priyadarshini, Amiya Bhusan Bagjadab, and Brojo Kishore Mishra

    Sushree Bibhuprada B. Priyadarshini, Amiya Bhusan Bagjadab, and Brojo Kishore Mishra. 2020. Natural Language Processing in Artifi- cial Intelligence. Apple Academic Press, Chapter A Brief Overview of Natural Language Processing and Artificial Intelligence, 211–224. doi:10.1201...

    2020 doi

  26. [34]

    Kian Long Tan, Chin Poo Lee, and Kian Ming Lim. 2023. A Survey of Sentiment Analysis: Approaches, Datasets, and Future Research. Applied Sciences 13, 7 (April 2023), 4550. doi:10.3390/app13074550

    2023 doi

  27. [35]

    Duyu Tang, Bing Qin, and Ting Liu. 2015. Document Modeling with Gated Recurrent Neural Network for Sentiment Classification. In 2015 Conference on Empirical Methods in Natural Language Process- ing (EMNLP). Association for Computational Linguistics, 1422–1432. doi:10.18653/v1/d15-1167

    2015 doi

  28. [36]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. At- tention is All you Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Sys- tem...

    2017 arXiv

  29. [37]

    Y. S. Weinstein, M. A. Pravia, E. M. Fortunato, S. Lloyd, and D. G. Cory

  30. [38]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S. Yu. 2021. A Comprehensive Survey on Graph Neural Networks. IEEE Transactions on Neural Networks and Learn- ing Systems 32, 1 (Jan. 2021), 4–24. doi: 10.1109/tnnls.2020.2978386 arXiv:1901.00596

    2021

  31. [39]

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyun- woo J. Kim. 2019. Graph Transformer Networks. In Advances in Neural Information Processing Systems 32: Annual Conference on Neu- ral Information Processing Systems NeurIPS 2019 . Curran Associates, Inc., 11960–11...

    2019 arXiv

  32. [40]

    Jonas Stein, Ivo Christ, Nicolas Kraus, Maximilian Balthasar Mansky, Robert Müller, and Claudia Linnhoff-Popien. 2023. Applying QNLP to Sentiment Analysis in Finance. In 2023 IEEE International Conference on Quantum Computing and Engineering (QCE) . IEEE, 20–25. doi:10. 1109/q...

    2023

  33. [48]

    Hui Zhang, Qinglin Zhao, and Chuangtao Chen. 2024. A light-weight quantum self-attention model for classical data classification. Applied Intelligence 54, 4 (Feb. 2024), 3077–3091. doi:10.1007/s10489-024-05337- w

    2024 doi

  34. [576]

    doi:10.1038/s43588-022-00311-3 arXiv:2303.09491

    arXiv

  35. [694]

    doi:10.1007/s11023-020-09548-1

    doi

  36. [2001]

    Physical Review Letters 86, 9 (Feb

    Implementation of the Quantum Fourier Transform. Physical Review Letters 86, 9 (Feb. 2001), 1889. doi:10.1103/physrevlett.86.1889 arXiv:quant-ph/9906059

    2001 arXiv

  37. [2014]

    In 2014 Conference on Empiri- cal Methods in Natural Language Processing (EMNLP)

    Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. In 2014 Conference on Empiri- cal Methods in Natural Language Processing (EMNLP) . Association for Computational Linguistics, 1724–1734. doi:10.3115/v1/d14-1179 arXiv:1406.1078

    2014 arXiv

  38. [2015]

    In 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    From Group to Individual Labels Using Deep Features. In 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery, 597–606. doi:10. 1145/2783258.2783380

  39. [2019]

    In 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long and Short Papers)

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long and Short Papers) . Association for Computational ...

    2019 arXiv

  40. [2023]

    In 24th Annual Conference of the European Association for Machine Translation

    Adaptive Machine Translation with Large Language Models. In 24th Annual Conference of the European Association for Machine Translation. European Association for Machine Translation, 227–237. arXiv:2301.13294 https://aclanthology.org/2023.eamt-1.22/

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.