Pith. sign in

REVIEW 5 major objections 5 minor 40 references

An Exploratory Evaluation of LLM-Assisted Rewriting of Moderate-Complexity Financial Sentences for DisCoCat-Based Sentiment Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read LLM-assisted rewriting can reduce DisCoCat circuit complexity for moderate-complexity financial sentences by more than 70 percent and raise observed mean accuracy over the low-complexity baseline, but the authors frame the result as an…

desk verdict Genuinely useful circuit-complexity measurements and an honest exploratory framing; the accuracy comparison, however, is not matched and should not be pushed as a headline result. read the letter →

arxiv 2608.07439 v1 pith:3HGBUIAE submitted 2026-08-07 cs.CL quant-ph

Brian Llinas , Nikos Chrisochoides This is my paper · ORCID
classification cs.CLquant-ph
keywords QuantumNaturalLanguageProcessingDisCoCatFinancialSentimentAnalysisLLM-assistedrewritingSemanticcompressionCircuitcomplexitylambeqPromptengineering
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 tests whether a large language model, used as a rewriting tool rather than a classifier, can turn moderate-complexity financial sentences into forms that a grammar-based quantum NLP pipeline can actually train on. The authors show that prompt-based compression cuts circuit size sharply: the strongest variant lowers average qubits by 75.81% and gates by 76.88% relative to the raw moderate-complexity subset. At the classification level, the best rewritten configuration reaches a mean accuracy of 0.550 ± 0.035, compared with 0.521 ± 0.050 for the low-complexity-only baseline. The paper is explicitly exploratory: it establishes feasibility and trade-offs, not a general accuracy improvement or an optimal preprocessing strategy.

What carries the argument

The carrying mechanism is the sentence-to-circuit compilation chain of DisCoCat, a compositional framework that turns grammatical structure into trainable quantum circuits: a sentence is parsed by BobcatParser into a pregroup grammar diagram, reduced, and mapped by lambeq into a parameterized circuit whose width (qubits) and gate count are determined by the grammatical structure. Rewriting intervenes before parsing, so shorter and syntactically simpler rewrites directly produce smaller circuits. The workflow's other load-bearing components are the three prompt strategies (Prompt A semantic compression, Prompt B parser-compatible rewriting, Prompt C decomposition) and the screening stage using MeaningBERT and FinBERT as proxies for meaning and sentiment preservation, with a threshold t ≥ 60 chosen as an exploratory operating point.

What would settle it

Take every rewrite that passes the t ≥ 60 MeaningBERT and FinBERT screens and have human annotators judge whether polarity, negation, and financial entities survive; if a substantial share of screened rewrites fail those judgments, the accuracy estimates and the claim that moderate-complexity inputs became usable collapse.

Watch

Extended reading notes

Core claim

The central discovery is that LLM-guided rewriting can serve as a preprocessing layer for DisCoCat-based financial sentiment analysis, converting longer 10-to-32-token synthetic sentences into shorter, parser-compatible forms that yield substantially smaller parameterized quantum circuits. The paper demonstrates this through three prompting strategies — semantic compression, parser-compatible rewriting, and decomposition — screened by MeaningBERT similarity, FinBERT sentiment agreement, and Bobcat parse success. Among successfully parsed outputs, GPT-4.1-mini with Prompt A achieves the largest circuit reductions, while GPT-4.1-mini with Prompt B achieves the highest observed mean accuracy. The authors state that the comparison with the Stein et al. baseline is descriptive, not a matched causal experiment, and that the results do not establish utility-scale deployment or a general accuracy improvement.

Load-bearing premise

The load-bearing premise is that MeaningBERT similarity and FinBERT sentiment agreement reliably capture whether a rewrite still carries the original sentiment-bearing meaning; the paper itself calls these proxies, not complete checks of propositions, negation, entities, or human faithfulness.

Editorial extensions

If this is right

  • Moderate-complexity sentences can be made trainable: GPT-4.1-mini with Prompt A lowers average qubits from 23.16 to 5.60 and gates from 118.07 to 27.29, a reduction of over 70 percent.
  • The best observed classifier, GPT-4.1-mini with Prompt B, reaches mean accuracy 0.550 ± 0.035 on an expanded 1.98× training set, compared with 0.521 ± 0.050 for the low-complexity-only reference.
  • Prompt C decomposition retains more meaning and more data under strict filtering but increases runtime roughly eightfold over the reference without improving accuracy, so data quantity alone does not drive DisCoCat performance.
  • Across configurations, training-split size shows a moderately negative Pearson correlation (r = -0.446) with accuracy, implying that how rewrites are selected matters more than how many are added.

Reading between the lines

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

  • If the proxy screens survive human evaluation, this rewriting layer could generalize beyond finance: any grammar-bound QNLP pipeline burdened by long sentences could compress inputs before parsing, at the cost of trusting LLM rewrites to preserve task-relevant meaning.
  • The depth reduction is much smaller than the qubit and gate reductions, so on real quantum hardware the practical speedup may be smaller than the simulation-side complexity numbers suggest; quantifying this would require backend or circuit-knitting experiments.
  • The negative split-size correlation is a hint that rewritten data may carry label noise or distribution shift; a clean test would fix prompt and filtering and vary only the number of rewritten instances added to the training set.
  • The paper's proposed tone/polarity-aware decomposition (Prompt D) is a natural next experiment: if selectively splitting only separable sentiment units beats both blanket compression and blanket decomposition, it would show that circuit-aware rewriting should be guided by sentiment structure, not just length.
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. This paper evaluates an LLM-assisted preprocessing workflow for DisCoCat-based financial sentiment analysis. Moderate-complexity synthetic financial sentences (mean 18.4 tokens; Table I) are rewritten via three prompting strategies (semantic compression, parser-compatible rewriting, and decomposition) using GPT-4.1-mini and Qwen2.5:7B, then screened for MeaningBERT similarity, FinBERT sentiment agreement, and Bobcat parse success before being merged with the low-complexity Stein et al. baseline and trained with lambeq DisCoCat. The main reported results are: (i) substantial circuit-level reductions, with GPT-4.1-mini + Prompt A lowering mean qubits by 75.81% and gates by 76.88% relative to the raw moderate subset (Table III); (ii) at the classification level, GPT-4.1-mini + Prompt B achieves the highest observed mean accuracy of 0.550±0.035 versus a 0.521±0.050 baseline (Table V); and (iii) a descriptive negative association between DisCoCat training-split size and accuracy (Pearson r=−0.446). The paper repeatedly and explicitly frames the accuracy comparisons as descriptive rather than matched, and its limitations section anticipates several of the concerns raised below.

Significance. The circuit-level half of the paper is a genuine contribution if it is taken at its stated scope. The reductions in Table III are concrete, internally consistent with the described lambeq/Bobcat pipeline, and arithmetically correct (5.60/23.16 gives a 75.81% qubit reduction; 27.29/118.07 gives a 76.88% gate reduction), and the code is publicly available, which materially strengthens reproducibility. The screening-cascade retention curves (Fig. 11) and the separation of compression versus decomposition mechanisms (Prompt A/B versus Prompt C) are informative for practitioners building grammar-aware QNLP preprocessing pipelines. The classification half is currently exploratory in the weakest sense: the headline 0.550 versus 0.521 comparison is not matched by test-set composition, is not significance-tested, the six-run standard deviations overlap, and filtered variants are subject to pre-split selection bias. As it stands, the paper demonstrates that moderate-complexity sentences can be parsed and compiled into much cheaper circuits, but it does not demonstrate that rewriting improves, or even enables, accurate classification of those sentences.

major comments (5)
  1. [III.D/III.F, Table V] The accuracy comparison is not matched, and the paper's 'descriptive' caveat does not repair the design gap. The augmented dataset is defined in Eq. (1) as D(v) = D_L ∪ A(v) and is split 80/10/10 after merging (Section III.F), so the test partition for every augmented variant mixes low-complexity baseline sentences with parser-valid rewritten moderate sentences, whereas the Stein et al. baseline of 0.521±0.050 was evaluated on low-complexity sentences only. No accuracy is reported for the rewritten-moderate subset alone anywhere in Section IV.C or Table V. Consequently, the observed gap of 0.550±0.035 versus 0.521±0.050 could be driven entirely by the low-complexity portion of the augmented test set, and the claim that moderate inputs are made 'usable' is not supported at the classification level by any reported number. The authors should report accuracy separately on the low-complexity test subset and on the rewritten-moderate test subset under identical training conditions, and should re-run the low-complexity-only condition in the same pipeline rather than borrowing the Stein et al. figure.
  2. [III.C/III.F, Table V] Filtering is applied before the data split, which biases the test partitions of the filtered variants. Strategy B (Section III.C) retains only outputs satisfying FinBERT label agreement and the MeaningBERT threshold before the merge in Eq. (1) and before the 80/10/10 split (Section III.F). For filtered variants, the test set therefore excludes moderate rewrites whose FinBERT prediction disagrees with the ground-truth label — the same labels against which downstream accuracy is measured — which can inflate achievable accuracy. The pattern in Table V is consistent with this concern: filtered GPT Prompt A (0.549±0.026) exceeds its unfiltered counterpart (0.502±0.012), although filtered GPT Prompt B (0.494±0.020) falls below its unfiltered counterpart (0.550±0.035), so the effect is configuration-dependent. Filtering must be applied inside the training split only, or the test set must be drawn from unfiltered data; the paper does not currently discuss this specific bias.
  3. [III.F, Table V] The headline accuracy advantage is within noise. Footnote 1 states that formal significance tests, confidence intervals, and effect-size analyses were not performed, and six repeated runs are used per configuration. With mean±SD values of 0.550±0.035 versus 0.521±0.050, the one-standard-deviation intervals overlap on [0.515, 0.571], so without a test the 'highest observed mean accuracy' claim carries no statistical weight even as a descriptive statement. A journal version should report per-configuration confidence intervals or a paired comparison across matched test sets; if the unmatched design precludes a meaningful test, the accuracy claims should be demoted from the abstract and from the list of contributions.
  4. [III.C, Fig. 11, Table V] The operating point t≥60 is selected post hoc from retention curves and is then used in the main filtered configurations of Table V (filtered GPT Prompt A and filtered GPT Prompt B). The paper correctly states that the threshold 'was not optimized using held-out data,' but that is precisely the problem: choosing t by inspecting the retained-sample curves of Fig. 11 and then reporting accuracy at that t as a headline result is a form of test-set-informed model selection. The authors should either fix t a priori or report downstream accuracy across t∈{40,50,60,70,80}; otherwise the filtered-row accuracies in Table V are conditional on a parameter chosen after seeing the data.
  5. [III.C, Section V] The claim that rewrites preserve sentiment-bearing meaning rests entirely on the MeaningBERT and FinBERT proxy scores, and Section V concedes that these are 'proxies rather than complete assessments of proposition preservation, negation, financial-entity preservation, or human-perceived faithfulness.' This is load-bearing because both the filtering cascade (Strategy B) and the downstream accuracy results are conditioned on these proxies: if the rewrites systematically distort sentiment in ways the proxies miss, the filtered training labels and the resulting accuracy estimates are invalid. The authors should add at minimum a human-judged faithfulness sample or a targeted negation/label-flip probe on the rewritten-moderate subset before claiming that those inputs are made usable for DisCoCat classification.
minor comments (5)
  1. [Section III.C, Fig. 3] The notation for the MeaningBERT threshold is inconsistent: the text defines s_i ≥ t, the workflow diagram in Fig. 3 labels the filter 'Meaning≥α,' and Fig. 11 uses t. Unify on a single symbol throughout.
  2. [Section III.F] Model versions, decoding settings, and software-library versions for the rewriting LLMs and for lambeq/pytket are not recorded, although the paper lists exactly this as future work; since a code link is provided and the Qwen Prompt A and Prompt C runs failed with a batching error, recording versions now would materially improve reproducibility.
  3. [Figure 5] The y-axis label 'Count' is confusing because the bars are not sample counts but the qubit, depth, and gate values for two example sentences; relabel the axes or present these numbers in the caption.
  4. [Figure 14] With only five or six runs per configuration, runs within a group are not independent observations, so the pooled Pearson r = −0.446 and Spearman ρ = −0.342 are at best descriptive; the paper says this, but the gray-point pooling in the figure invites over-reading.
  5. [Section II.C] Reference [28] is cited in the text as 'Juseod-DO et al.'; the author name should be given as 'Do et al.' (Juseon Do), matching the reference list entry.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical pipeline study whose reported circuit and accuracy quantities are measured, not derived, and whose only self-citation is not load-bearing.

full rationale

This is an empirical evaluation rather than a derivation, so the derivation-equivalence patterns do not apply. The reported quantities—circuit complexity reductions, parser validity rates, retained dataset sizes, and observed classification accuracies—are measured from executed pipelines (Bobcat parsing, lambeq circuit construction, and six repeated training runs), not computed from fitted parameters that are then relabeled as predictions. The paper explicitly frames the Stein et al. comparison as descriptive and 'not a fully matched causal experiment,' which is a validity limitation but not circularity. The MeaningBERT threshold t>=60 is selected from retention curves and is repeatedly described as an exploratory operating point that was not optimized on held-out data; this is post hoc selection, not a self-definitional or fitted-input circle. The only self-citation involving an author of the present paper (Stenger et al. [39], co-authored by Chrisochoides) appears in a future-work sentence about circuit knitting and does not support any central claim or reported result. No equation in the paper defines a predicted outcome in terms of the very result it is used to establish, and no fitted input is renamed as a prediction. The paper is therefore self-contained with respect to circularity concerns, and its acknowledged limitations belong to experimental design and external validity rather than circular reasoning.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a modest set of empirical choices rather than a derivation. The main free parameters are the screening threshold and prompt designs. The paper introduces no new theoretical entities. The key assumptions are the validity of the synthetic data and the proxy screening measures, both of which the authors disclose as limitations.

free parameters (3)
  • MeaningBERT threshold t = 60 (exploratory)
    Selected as an operating point after inspecting retention curves, not predefined or optimized on held-out data. Affects the retained training set and therefore all downstream accuracy numbers.
  • Prompt designs A, B, C = Hand-authored
    The three prompt strategies were designed by the authors. The exact wording is not fully specified, and prompt design is a free choice that directly influences rewriting outcomes.
  • FinBERT label-agreement filter = Filter requires predicted label = ground truth
    An inclusion criterion for Strategy B that shapes which rewrites enter training. The combination of this filter with the MeaningBERT threshold is a joint screening choice with no predefined operating rule.
assumptions (4)
  • domain assumption The synthetic ChatGPT-generated financial sentiment subsets from Stein et al. are a valid testbed for financial sentiment analysis.
    The paper builds entirely on these synthetic subsets and states they do not capture the diversity and noise of natural financial text. The claim of 'usability' is only about this synthetic distribution.
  • domain assumption MeaningBERT and FinBERT scores are acceptable proxies for semantic and sentiment preservation.
    Used as screening signals and as evidence that rewrites preserve sentiment. The paper explicitly acknowledges they are proxies, not complete assessments of proposition preservation, negation, or financial-entity preservation.
  • domain assumption The lambeq/Bobcat pipeline correctly maps grammar to quantum circuits, so qubit and gate counts reflect computational complexity.
    The circuit-level analysis assumes the DisCoCat compilation through Bobcat and IQPAnsatz is a faithful and relevant complexity measure. This is standard in the QNLP literature but still a framework assumption.
  • domain assumption Accuracy is the appropriate evaluation metric for comparing DisCoCat sentiment variants.
    The paper evaluates only accuracy, noting it does not fully characterize model behavior. This choice is consistent with the prior Stein et al. baseline but narrow.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Exploratory Evaluation of LLM-Assisted Rewriting of Moderate-Complexity Financial Sentences for DisCoCat-Based Sentiment Analysis." pith.science (2026). https://pith.science/paper/3HGBUIAE

@misc{pith2026260807439,
  author       = {Pith},
  title        = {Pith review of: An Exploratory Evaluation of LLM-Assisted Rewriting of Moderate-Complexity Financial Sentences for DisCoCat-Based Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3HGBUIAE}},
  note         = {Machine review of arXiv:2608.07439}
}
abstract

Quantum natural language processing (QNLP) provides a grammar-aware framework for text modeling, and Distributional Compositional Categorical (DisCoCat) is one of its theoretically grounded formulations. Prior work on financial sentiment analysis has identified practical limitations of DisCoCat, including parser sensitivity, high simulation cost, and difficulty handling longer sentences. We study an LLM-assisted preprocessing workflow that uses controlled rewriting to compress, simplify, or decompose moderate-complexity financial sentiment sentences into parser-compatible, circuit-efficient variants while preserving sentiment-bearing meaning. We compare prompting strategies, language models, and filtering configurations with the low-complexity-only DisCoCat baseline of Stein et al. At the circuit level, the strongest compression variants reduce average qubit and gate counts by more than 70 percent relative to the raw moderate-complexity subset. Across repeated training runs, GPT-4.1-mini with Prompt B achieves the highest observed mean accuracy, $0.550 \pm 0.035$, compared with $0.521 \pm 0.050$ for the baseline. Larger training splits do not necessarily improve downstream performance; across evaluated configurations, training-split size has a moderately negative association with accuracy (Pearson $r=-0.446$). These results provide exploratory evidence that LLM-assisted rewriting can make some moderate-complexity inputs usable within the evaluated DisCoCat configuration, while highlighting prompt design, filtering, and circuit-aware preprocessing as considerations for more scalable QNLP-based financial sentiment analysis.

Figures

Figures reproduced from arXiv: 2608.07439 by the authors.

Figure 1
Figure 1. First two stages of the compilation process for [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Final stage of the compilation process for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Exploratory workflow for generating, screening, filtering, and evaluating LLM-assisted rewrites of moderate-complexity synthetic financial sentences [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Illustrative comparison of circuit complexity for one low-complexity [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Token-count distributions after prompt-based rewriting of the raw moderate-complexity subset. Prompt A and Prompt B strongly compress sentence [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Corpus-level circuit complexity across rewriting variants. From left to right, the boxplots show the distributions of qubit count, circuit depth, and gate [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Distribution of MeaningBERT similarity scores for rewritten outputs. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Automated FinBERT sentiment-fidelity diagnostics for rewritten [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Bobcat parser success rates for rewritten outputs at the row and [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Training-set retention under progressively stricter screening conditions. Panel (a) applies the MeaningBERT threshold only; panel (b) additionally [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Run-level distribution of final test accuracy and runtime across configurations. Boxplots summarize repeated runs, and overlaid points indicate [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Epoch-level distributions of training and validation accuracy and loss across evaluated configurations. Boxplots summarize values pooled across [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Observed final test accuracy across DisCoCat training-split-size groups. Gray points represent individual runs, while blue markers and error bars [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 36 canonical work pages

  1. [1]

    Combining Technical Analysis with Sentiment Analysis for Stock Price Prediction,

    S. Deng, T. Mitsubuchi, K. Shioda, T. Shimada, and A. Sakurai, “Combining Technical Analysis with Sentiment Analysis for Stock Price Prediction,” in2011 IEEE Ninth International Conference on Dependable, Autonomic and Secure Computing, 2011, pp. 800–807

    work page 2011

  2. [2]

    Invest- ment strategies applied to the Brazilian stock market: A methodology based on Sentiment Analysis with deep learning,

    A. E. de Oliveira Carosia, G. P. Coelho, and A. E. A. da Silva, “Invest- ment strategies applied to the Brazilian stock market: A methodology based on Sentiment Analysis with deep learning,”Expert Systems with Applications, vol. 184, p. 115470, 2021

    work page 2021

  3. [3]

    Sentiment Analysis Based on Financial Tweets and Market Information,

    S. Ao, “Sentiment Analysis Based on Financial Tweets and Market Information,” in2018 International Conference on Audio, Language and Image Processing (ICALIP), 2018, pp. 321–326

    work page 2018

  4. [4]

    LSTM-based sentiment analysis for stock price forecast,

    C.-R. Ko and H.-T. Chang, “LSTM-based sentiment analysis for stock price forecast,”PeerJ Computer Science, vol. 7, p. e408, 2021

    work page 2021

  5. [5]

    Transformers in the Real World: A Survey on NLP Applications,

    N. Patwardhan, S. Marrone, and C. Sansone, “Transformers in the Real World: A Survey on NLP Applications,”Information, vol. 14, p. 242,

    work page

  6. [6]

    A Review on Stock Market Trends and Stocks Price Prediction Using Sentiment Analysis and Market Data,

    N. Mishra, M. Anzar, S. Pandey, and S. Mishra, “A Review on Stock Market Trends and Stocks Price Prediction Using Sentiment Analysis and Market Data,” in2025 3rd International Conference on Communi- cation, Security, and Artificial Intelligence (ICCSAI), vol. 3, 2025, pp. 56–62

    work page 2025

  7. [7]

    Natural language processing in finance: A survey,

    K. Du, Y . Zhao, R. Mao, F. Xing, and E. Cambria, “Natural language processing in finance: A survey,”Information Fusion, vol. 115, p. 102755, 2025

    work page 2025

  8. [8]

    Text-based sentiment analysis in finance: Synthesising the existing literature and exploring future directions,

    A. Todd, J. Bowden, and Y . Moshfeghi, “Text-based sentiment analysis in finance: Synthesising the existing literature and exploring future directions,”Intelligent Systems in Accounting, Finance and Management, vol. 31, p. e1549, 2024

    work page 2024

Show all 40 references
  1. [9]

    Evaluation of Sentiment Analysis in Finance: From Lexicons to Transformers,

    K. Mishev, A. Gjorgjevikj, I. V odenska, L. T. Chitkushev, and D. Tra- janov, “Evaluation of Sentiment Analysis in Finance: From Lexicons to Transformers,”IEEE Access, vol. 8, 2020

    2020

  2. [10]

    New mean- ing for NLP: the trials and tribulations of natural language processing with GPT-3 in ophthalmology,

    S. Nath, A. Marie, S. Ellershaw, E. Korot, and P. A. Keane, “New mean- ing for NLP: the trials and tribulations of natural language processing with GPT-3 in ophthalmology,”British Journal of Ophthalmology, vol. 106, pp. 889–892, 2022

    2022

  3. [11]

    Efficient large-scale language model training on GPU clusters using megatron-LM,

    D. Narayanan, M. Shoeybi, J. Casper, P. LeGresley, M. Patwary, V . Korthikanti, D. Vainbrand, P. Kashinkunti, J. Bernauer, B. Catanzaro, A. Phanishayee, and M. Zaharia, “Efficient large-scale language model training on GPU clusters using megatron-LM,” inProceedings of the Inte...

    2021

  4. [12]

    Generalization in quantum machine learning from few training data,

    M. C. Caro, H.-Y . Huang, M. Cerezo, K. Sharma, A. Sornborger, L. Cincio, and P. J. Coles, “Generalization in quantum machine learning from few training data,”Nature Communications, vol. 13, p. 4919, 2022

    2022

  5. [13]

    Quantum Natural Language Processing: Challenges and Opportunities,

    R. Guarasci, G. De Pietro, and M. Esposito, “Quantum Natural Language Processing: Challenges and Opportunities,”Applied Sciences, vol. 12, p. 5651, 2022

    2022

  6. [14]

    Quantum Natural Language Processing on Near-Term Quantum Computers,

    K. Meichanetzidis, S. Gogioso, G. d. Felice, N. Chiappori, A. Toumi, and B. Coecke, “Quantum Natural Language Processing on Near-Term Quantum Computers,”Electronic Proceedings in Theoretical Computer Science, vol. 340, pp. 213–229, 2021

    2021

  7. [15]

    QNLP in Practice: Running Compositional Models of Meaning on a Quantum Computer,

    R. Lorenz, A. Pearson, K. Meichanetzidis, D. Kartsaklis, and B. Coecke, “QNLP in Practice: Running Compositional Models of Meaning on a Quantum Computer,”Journal of Artificial Intelligence Research, vol. 76, pp. 1305–1342, 2023

    2023

  8. [16]

    Towards Quantum Advantage in Financial Market Risk using Quantum Gradient Algorithms,

    N. Stamatopoulos, G. Mazzola, S. Woerner, and W. J. Zeng, “Towards Quantum Advantage in Financial Market Risk using Quantum Gradient Algorithms,”Quantum, vol. 6, p. 770, 2022

    2022

  9. [17]

    The Dawn of Quantum Natural Language Processing,

    R. Di Sipio, J.-H. Huang, S. Y .-C. Chen, S. Mangini, and M. Worring, “The Dawn of Quantum Natural Language Processing,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 8612–8616, iSSN: 2379-190X

    2022

  10. [18]

    Quantum Long Short-Term Memory,

    S. Y .-C. Chen, S. Yoo, and Y .-L. L. Fang, “Quantum Long Short-Term Memory,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 8622– 8626

    2022

  11. [19]

    A multiclass Q-NLP sentiment analysis experiment using DisCoCat,

    V . Martinez and G. Leroy-Meline, “A multiclass Q-NLP sentiment analysis experiment using DisCoCat,” 2022, arXiv:2209.03152 [cs]

    2022 arXiv

  12. [20]

    Applying QNLP to Sentiment Analysis in Finance,

    J. Stein, I. Christ, N. Kraus, M. B. Mansky, R. M ¨uller, and C. Linnhoff- Popien, “Applying QNLP to Sentiment Analysis in Finance,” in2023 IEEE International Conference on Quantum Computing and Engineer- ing (QCE), vol. 02, 2023, pp. 20–25

    2023

  13. [21]

    Extending Context Window of Large Language Models via Semantic Compression,

    W. Fei, X. Niu, P. Zhou, L. Hou, B. Bai, L. Deng, and W. Han, “Extending Context Window of Large Language Models via Semantic Compression,” inFindings of the Association for Computational Lin- guistics: ACL 2024, 2024, pp. 5169–5181

    2024

  14. [22]

    Mathematical Foundations for a Compositional Distributional Model of Meaning,

    B. Coecke, M. Sadrzadeh, and S. Clark, “Mathematical Foundations for a Compositional Distributional Model of Meaning,” Mar. 2010, arXiv:1003.4394 [cs]. [Online]. Available: http://arxiv.org/abs/1003. 4394

    2010 arXiv

  15. [23]

    Financial Sentiment Analysis: Techniques and Applications,

    K. Du, F. Xing, R. Mao, and E. Cambria, “Financial Sentiment Analysis: Techniques and Applications,”ACM Comput. Surv., vol. 56, pp. 220:1– 220:42, 2024

    2024

  16. [24]

    Simple Sentiment Analysis Ansatz for Sentiment Clas- sification in Quantum Natural Language Processing,

    F. Z. Ruskanda, M. R. Abiwardani, I. Syafalni, H. T. Larasati, and R. Mulyawan, “Simple Sentiment Analysis Ansatz for Sentiment Clas- sification in Quantum Natural Language Processing,”IEEE Access, vol. 11, pp. 120 612–120 627, 2023

    2023

  17. [25]

    An Effective Strategy for Sentiment Analysis Based on Complex-Valued Embedding and Quantum Long Short-Term Memory Neural Network,

    Z. Chu, X. Wang, M. Jin, N. Zhang, Q. Gao, and L. Shao, “An Effective Strategy for Sentiment Analysis Based on Complex-Valued Embedding and Quantum Long Short-Term Memory Neural Network,”Axioms, vol. 13, p. 207, 2024

    2024

  18. [26]

    Semantic Compression with Large Language Models,

    H. Gilbert, M. Sandborn, D. C. Schmidt, J. Spencer-Smith, and J. White, “Semantic Compression with Large Language Models,” in2023 Tenth International Conference on Social Networks Analysis, Management and Security (SNAMS), 2023, pp. 1–8, iSSN: 2831-7343

    2023

  19. [27]

    Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference,

    B. Liskavets, M. Ushakov, S. Roy, M. Klibanov, A. Etemad, and S. K. Luke, “Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, pp. 24 595–24 604, 2025

    2025

  20. [28]

    InstructCMP: Length Control in Sentence Compression through Instruction-based Large Language Models,

    Juseon-Do, J. Kwon, H. Kamigaito, and M. Okumura, “InstructCMP: Length Control in Sentence Compression through Instruction-based Large Language Models,” inFindings of the Association for Compu- tational Linguistics: ACL 2024, 2024, pp. 8980–8996

    2024

  21. [29]

    Redefining Simplicity: Benchmarking Large Language Models from Lexical to Document Simplification,

    J. Qiang, M. Huang, Y . Zhu, Y . Yuan, C. Zhang, and K. Yu, “Redefining Simplicity: Benchmarking Large Language Models from Lexical to Document Simplification,” 2025, arXiv:2502.08281 [cs]

    2025 arXiv

  22. [30]

    LLM-based Text Simplification and its Effect on User Comprehension and Cognitive Load,

    T. Guidroz, D. Ardila, J. Li, A. Mansour, P. Jhun, N. Gonzalez, X. Ji, M. Sanchez, S. Kakarmath, M. M. Bellaiche, M. ´A. Garrido, F. Ahmed, D. Choudhary, J. Hartford, C. Xu, H. J. S. Echeverria, Y . Wang, J. Shaffer, Eric, Cao, Y . Matias, A. Hassidim, D. R. Webster, Y . Liu, ...

    2025 arXiv

  23. [31]

    Language Models are Few-Shot Learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert- V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. C...

    2020

  24. [32]

    Prompt Engineering as an Important Emerging Skill for Medical Professionals: Tutorial,

    B. Mesk ´o, “ Prompt Engineering as an Important Emerging Skill for Medical Professionals: Tutorial,”Journal of Medical Internet Research, vol. 25, p. e50638, 2023

    2023

  25. [33]

    Large Language Models are Zero-Shot Reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large Language Models are Zero-Shot Reasoners,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 22 199–22 213

    2022

  26. [34]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhouet al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 24 824–24 837

    2022

  27. [35]

    Prompt Engineering with ChatGPT: A Guide for Academic Writers,

    L. Giray, “Prompt Engineering with ChatGPT: A Guide for Academic Writers,”Annals of Biomedical Engineering, vol. 51, no. 12, pp. 2629– 2633, 2023

    2023

  28. [36]

    How to write effective prompts for large language models,

    Z. Lin, “How to write effective prompts for large language models,” Nature Human Behaviour, vol. 8, no. 4, pp. 611–615, 2024

    2024

  29. [37]

    MeaningBERT: assessing meaning preservation between sentences,

    D. Beauchemin, H. Saggion, and R. Khoury, “MeaningBERT: assessing meaning preservation between sentences,”Frontiers in Artificial Intelli- gence, vol. 6, 2023

    2023

  30. [38]

    FinBERT: Financial Sentiment Analysis with Pre-trained Language Models,

    D. Araci, “FinBERT: Financial Sentiment Analysis with Pre-trained Language Models,” 2019, arXiv:1908.10063 [cs]

    2019 arXiv

  31. [39]

    Scalable quantum circuit knitting using a weak-coupling approximation,

    J. P. T. Stenger, D. Gunlycke, and N. Chrisochoides, “Scalable quantum circuit knitting using a weak-coupling approximation,” Jun. 2026, arXiv:2606.19035 [quant-ph]. [Online]. Available: http: //arxiv.org/abs/2606.19035

    2026 arXiv

  32. [2023]

    Available: https://www.mdpi.com/2078-2489/14/4/242

    [Online]. Available: https://www.mdpi.com/2078-2489/14/4/242

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.