Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Semantic-preserved Augmentation with Confidence-weighted Fine-tuning for Aspect Category Sentiment Analysis

T0 review · 4 major / 4 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 LLM paraphrase augmentation, semantic filtering, and confidence-weighted fine-tuning achieve the best F1 scores on four aspect-category sentiment benchmarks.

desk verdict The headline gains are almost certainly model scale, not the proposed augmentation or loss; the paper's own Table IV shows it. read the letter →

arxiv 2506.07148 v1 pith:QCRJ7JKR submitted 2025-06-08 cs.CL

Yaping Chai , Haoran Xie , Joe S. Qin This is my paper
classification cs.CL
keywords aspectcategorysentimentanalysisdataaugmentationlargelanguagemodelssemanticconsistencyfilteringconfidence-weightedfine-tuningSentence-BERTFlan-T5low-resource
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 a three-stage pipeline solves data scarcity in aspect category sentiment analysis (ACSA), the task of jointly detecting which predefined aspect categories a sentence mentions and the sentiment attached to each. Its core claim is that LLM-generated paraphrases that preserve the original sentence's meaning, filtered by Sentence-BERT cosine similarity, expand the training distribution without corrupting label pairs, and that reweighting the cross-entropy loss by the model's own confidence on correct predictions yields more confident and accurate outputs. On the Rest15, Rest16, Lap15, and Lap16 benchmarks, the authors report that their method, built on Flan-T5-xl, beats every cited baseline for both aspect-category sentiment classification and the end-to-end variant. If the paper is right, it provides a low-cost, label-preserving route to stronger fine-grained sentiment models where annotated data is scarce.

What carries the argument

The argument rides on three components: a structured prompt template that asks the LLM to paraphrase while holding the category-polarity pairs fixed; a Sentence-BERT semantic consistency filter that retains only synthetic sentences whose cosine similarity to the original is at least $\tau$; and the confidence-weighted loss $L^v_i = L_{\mathrm{CE}}(y_i)(1 + \alpha v_i)$, in which $v_i$ is the model's maximum predicted probability when the prediction is correct and zero otherwise. The template ensures the generated data inherits the original labels, the filter removes semantic drift, and the loss amplifies gradients for correct, confident predictions while leaving incorrect predictions on standard cross-entropy.

What would settle it

Fine-tune the same Flan-T5-xl on the original four training sets with standard cross-entropy and no augmentation, and compare its F1 with the paper's w/o DA and DA rows; if it matches Ours (w/o DA) on Rest16 and Lap16 ACSA, the central claim that augmentation and confidence weighting drive the gains is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is a generative augmentation and training recipe for aspect category sentiment analysis. It prompts GPT-4o with a structured template that carries the original sentence's aspect-category and sentiment pairs, forcing generated rewrites to keep the same categories and polarities while the surface wording changes; a post-processing step then discards any synthetic sentence whose Sentence-BERT cosine similarity to the original falls below an empirically set threshold of $\tau = 0.7$. On the training side, the authors replace plain cross-entropy with a confidence-weighted loss $L^v_i = L_{\mathrm{CE}}(y_i)(1 + \alpha v_i)$, where $v_i$ is the model's maximum class probability when the prediction is correct and zero otherwise. They report that this method, fine-tuned on Flan-T5-xl, obtains the best F1 among all compared methods on Rest15, Rest16, Lap15, and Lap16 for both ACSC and ACSA; for example, ACSC F1 reaches 93.44 on Rest16 against 82.20 for the strongest baseline, and ACSA F1 reaches 82.65 on Rest16 against 75.03. On some datasets the unaugmented variant is better than the augmented one, so the paper attributes part of the gain to the base model and the confidence-weighted fine-tuning rather than to augmentation alone.

Load-bearing premise

The head-to-head comparisons assume that a 3-billion-parameter instruction-tuned Flan-T5-xl against BERT-scale baselines is a fair test of the proposed augmentation and loss; if a plain Flan-T5-xl without these components matches the reported scores, the advantage comes from model scale rather than the method.

Editorial extensions

If this is right

  • Practitioners can expand ACSA training sets without manual relabeling, because the prompt template forces generated paraphrases to keep the original aspect categories and sentiment polarities.
  • Filtering synthetic sentences by semantic similarity improves downstream performance on most of the tested datasets, indicating that data quality control can matter more than generation volume.
  • Confidence-weighted fine-tuning alone, without augmentation, already exceeds the cited BERT-scale baselines, suggesting the loss helps even in the absence of extra data.
  • On domains where augmentation slightly hurts performance, keeping the unaugmented training scheme avoids degradation, so augmentation is dataset-dependent rather than universally beneficial.
  • The reported gains are consistent across restaurant and laptop domains, supporting the method's generality for fine-grained category-sentiment tasks.

Reading between the lines

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

  • The headline comparison does not isolate model scale: the paper's own model-size sweep shows Rest16 ACSA F1 jumping from 56.12 with a 220M model to 82.65 with a 3B model, so a large share of the gap over BERT-scale baselines may come from Flan-T5-xl's capacity and instruction tuning rather than from augmentation or the confidence-weighted loss.
  • A fair component test would run the same Flan-T5-xl with standard cross-entropy and no augmentation, then add each proposed component separately at each model size; the paper does not report that ablation.
  • The confidence-weighted loss rewards already-correct high-confidence predictions, which could push the model toward overconfidence; the per-dataset tuning of $\alpha$ via Optuna leaves open whether a single default $\alpha$ would retain the reported advantage.
  • A natural extension is to apply the structured prompt and SBERT filter to other ABSA subtasks, such as aspect term extraction, where label preservation across paraphrases could be tested directly.
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

4 major / 4 minor

Summary. The paper proposes a pipeline for aspect category sentiment analysis (ACSA/ACSC) consisting of LLM-generated paraphrases with a structured prompt, Sentence-BERT semantic-similarity filtering, and a confidence-weighted cross-entropy fine-tuning loss. It evaluates on Rest15, Rest16, Lap15, and Lap16 using Flan-T5-xl as the backbone and reports F1 scores that are higher than several published BERT-scale baselines. The paper also includes analyses of the confidence weight alpha, the similarity threshold tau, model scale, and error distributions. The central claim is that the proposed augmentation and confidence-weighted fine-tuning consistently achieve state-of-the-art performance on all four datasets.

Significance. If the claims were properly supported, the paper would offer a simple and potentially practical recipe for low-resource ACSA: a prompt template for LLM paraphrase generation, an SBERT filter, and a modified training objective. The paper has useful components, including a concrete prompt template, a parameter analysis, and a category-level error analysis that identifies interpretable failure patterns. However, the empirical support for the central claim is currently weak. The headline comparisons pit a 3B-parameter instruction-tuned model against BERT-scale baselines whose numbers are imported from prior papers, the contribution of the confidence-weighted loss is not isolated by any ablation, and the augmentation itself is inconsistent, sometimes lowering F1 relative to the no-augmentation variant. The significance of the work is therefore conditional on a substantially strengthened experimental design.

major comments (4)
  1. [IV.D3, Table IV] The model-size study exposes a scale confound that undermines the headline SOTA claim. On Rest16 ACSA, the proposed method achieves F1 56.12 at 220M, 64.12 at 770M, and 82.65 at 3B, while the PBJM baseline from Table III reports 75.03. Since all baseline numbers in Tables II and III are from BERT-scale models and are copied from ECAN and PBJM, the observed gaps at 3B cannot be attributed to the proposed augmentation or confidence-weighted loss rather than to model scale and instruction tuning. The authors should report the critical control: Flan-T5-xl fine-tuned with plain cross-entropy, without augmentation and without confidence weighting, on all four datasets and both tasks. Without this matched-scale baseline, the abstract's claim of consistent state-of-the-art performance is not established.
  2. [III.C, Eqs. (2)-(4)] No experiment isolates the confidence-weighted loss. The two reported variants, Ours (DA) and Ours (w/o DA), presumably both use the confidence-weighted objective, but there is no row corresponding to standard cross-entropy fine-tuning of the same Flan-T5-xl model with the same data. Consequently, the improvement attributed to Eq. (3) is untested. The authors should add ablations with and without the confidence weighting, with and without augmentation, so that each component's contribution can be evaluated separately.
  3. [IV.C, Tables II and III] The data augmentation component is not consistently beneficial, which contradicts the paper's framing. On Rest15 ACSC, Ours (DA) obtains F1 86.72 versus Ours (w/o DA) 89.46; on Rest15 ACSA the corresponding numbers are 74.48 versus 77.44; and on Lap16 ACSA they are 60.45 versus 61.56. The paper acknowledges these decreases but still claims that the proposed method, which includes augmentation, consistently achieves the best performance. The authors should either restrict the claim to the configurations that actually benefit, or provide evidence, such as multiple seeds and significance tests, that the augmentation helps on average. As written, the augmentation cannot be considered a validated component of the method.
  4. [IV.B and IV.D1] The selection of the hyperparameter alpha is reported inconsistently and without a clear validation protocol. Section IV.B lists alpha values tuned with Optuna, e.g., 0.5903 for Rest16 ACSA and 0.5378 for Lap16 ACSA, while Section IV.D1 reports that the best Rest16 ACSA performance is at alpha=1.0 and the best Lap16 ACSA performance at alpha=0.4. Similarly, Rest15 ACSA is listed as 0.1713 but Figure 4 is said to peak at alpha=0.2. If alpha was selected using test-set performance, the reported results are optimistically biased. The authors need to specify the validation split, the selection rule, and reconcile the alpha values between the main tables and the parameter analysis.
minor comments (4)
  1. [IV.E3] There is a typo in the category name 'RESITAURANT#GENERAL'; it should be 'RESTAURANT#GENERAL'.
  2. [III.B1] The sentence 'Pair Sgen generated by LLM with the original label y to form an augmented sample' is grammatically incomplete; it should read 'The synthetic sentence Sgen is paired with the original label y to form an augmented sample x′=(Sgen,y).'
  3. [IV.D2] The text says that tau=0.7 provides the best trade-off across all datasets, but the preceding paragraph states that on Rest15, tau=0 performs best. This should be stated more precisely, for example by noting that tau=0.7 is best on the other three datasets.
  4. [IV.A] The paper does not mention whether results are averaged over multiple random seeds or whether a fixed seed was used. Given the small F1 differences in some comparisons, reporting variance or significance would strengthen the conclusions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed augmentation and confidence-weighted fine-tuning are empirically evaluated against external benchmarks, and no claimed result reduces by construction to its inputs.

full rationale

The paper's central claims are empirical: GPT-4o paraphrase augmentation with SBERT filtering and confidence-weighted fine-tuning on Flan-T5-xl is reported to outperform baselines on four public SemEval benchmarks (Tables II and III). The derivation chain is not circular. The augmented sentences are generated from original sentences plus their labels, filtered by cosine similarity using a pretrained external SBERT model, and then used with the original labels in standard supervised fine-tuning. The confidence-weighted loss (Eqs. 2-3) reweights the cross-entropy loss by the model's own predicted probability on correct predictions; while this is self-referential within training, it is a training heuristic rather than a logical derivation in which the output is defined as the input. No 'prediction' is fitted and then renamed as a result: the reported F1 scores are measured against held-out test sets from SemEval, and the most competitive baselines' numbers are copied from ECAN and PBJM rather than derived from the proposed method. The hyperparameters alpha and tau are tuned empirically, which raises standard model-selection and comparison-fairness concerns, but tuning a loss weight does not make the final benchmark score equivalent to the tuned parameter by construction. The authors' self-citations ([1], [21], [22]) are contextual or survey references and are not load-bearing: removing them would not change the method or the empirical evaluation. The model-scale confound identified by a skeptical reader (Table IV: 220M vs 3B Flan-T5) is a genuine validity threat to attributing the gains to the proposed components, but it is a correctness or experimental-design issue, not circularity. There is no self-citation chain invoked to forbid alternatives, no uniqueness theorem imported from the authors' prior work, and no known result merely renamed. Accordingly, the appropriate circularity score is 0.

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

The central empirical claim rests on several unverified assumptions: paraphrases preserve labels, the SBERT threshold is a valid label-preservation check, and the same-scale base model without the proposed loss would not already match the results. The loss introduces per-dataset hyperparameters (alpha, tau) that are tuned without a documented validation protocol.

free parameters (3)
  • alpha (confidence weight) = ACSC: 0.4673, 0.3045, 0.3045, 0.6123; ACSA: 0.1713, 0.5903, 0.2526, 0.5378
    Tuned per dataset and task with Optuna; validation split not specified, and Figure 4 plots test metrics against alpha, indicating possible test-set fitting.
  • tau (semantic similarity threshold) = 0.7
    Determined empirically; on Rest15, tau=0 (no filtering) gives the best F1, so the global 0.7 does not dominate all datasets.
  • number of generated sentences per training sample
    Not reported; without this, the augmentation budget is unspecified.
assumptions (4)
  • domain assumption GPT-4o paraphrases with the original labels y retain the true category-sentiment pairs.
    Invoked in Section III.B.1 when forming x' = (Sgen, y); a violation would inject mislabeled training data.
  • domain assumption SBERT cosine similarity above 0.7 is a valid proxy for label-preserving semantic equivalence.
    Used in Section III.B.2 to accept or reject generated sentences; no evidence links this threshold to label correctness.
  • domain assumption Flan-T5-xl fine-tuned with standard cross-entropy loss would not already achieve the reported gains.
    The paper never ablates standard fine-tuning of the same model, so this assumption underlies attribution of the gains to the proposed loss.
  • domain assumption The cited baseline results (ECAN, PBJM) use the same evaluation protocol and data splits.
    All baseline numbers are taken from the respective papers; no re-implementation or protocol check is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic-preserved Augmentation with Confidence-weighted Fine-tuning for Aspect Category Sentiment Analysis." pith.science (2026). https://pith.science/paper/QCRJ7JKR

@misc{pith2026250607148,
  author       = {Pith},
  title        = {Pith review of: Semantic-preserved Augmentation with Confidence-weighted Fine-tuning for Aspect Category Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QCRJ7JKR}},
  note         = {Machine review of arXiv:2506.07148}
}
read the original abstract

Large language model (LLM) is an effective approach to addressing data scarcity in low-resource scenarios. Recent existing research designs hand-crafted prompts to guide LLM for data augmentation. We introduce a data augmentation strategy for the aspect category sentiment analysis (ACSA) task that preserves the original sentence semantics and has linguistic diversity, specifically by providing a structured prompt template for an LLM to generate predefined content. In addition, we employ a post-processing technique to further ensure semantic consistency between the generated sentence and the original sentence. The augmented data increases the semantic coverage of the training distribution, enabling the model better to understand the relationship between aspect categories and sentiment polarities, enhancing its inference capabilities. Furthermore, we propose a confidence-weighted fine-tuning strategy to encourage the model to generate more confident and accurate sentiment polarity predictions. Compared with powerful and recent works, our method consistently achieves the best performance on four benchmark datasets over all baselines.

Figures

Figures reproduced from arXiv: 2506.07148 by the authors.

Figure 1
Figure 1. An illustration of ACD, ACSC and ACSA subtasks. The c, s, and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overview of our method, including data augmen [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Structured prompt template for generating sentences. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Precision, Recall, and F1 scores on four datasets with different values of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Precision, Recall, and F1 scores on four datasets with different values of the semantic consistency threshold [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Category error distribution and Sentiment error distri [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Error counts for mispredicted aspect categories across the four benchmark datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 6
Figure 6. Figure 6: 1) Identifying the correct category is more difficult than identifying its sentiment.: Category errors exceed Sentiment errors in all benchmark datasets, and for the Restaurant dataset, Category errors are almost three times that of Sentiment errors. For the Laptop dat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 31 canonical work pages

  1. [1]

    Cognitive- inspired deep learning models for aspect-based sentiment analysis: A retrospective overview and bibliometric analysis,

    X. Chen, H. Xie, S. J. Qin, Y . Chai, X. Tao, and F. L. Wang, “Cognitive- inspired deep learning models for aspect-based sentiment analysis: A retrospective overview and bibliometric analysis,”Cogn. Comput., vol. 16, no. 6, pp. 3518–3556, 2024

    2024

  2. [2]

    Exploring cognitive and aesthetic causality for multimodal aspect-based sentiment analysis,

    L. Xiao, R. Mao, S. Zhao, Q. Lin, Y . Jia, L. He, and E. Cambria, “Exploring cognitive and aesthetic causality for multimodal aspect-based sentiment analysis,” IEEE Trans. Affective Comput. , pp. 1–18, 2025

    work page 2025

  3. [3]

    Mer 2025: When affective computing meets large language models,

    Z. Lian, R. Liu, K. Xu, B. Liu, X. Liu, Y . Zhang, X. Liu, Y . Li, Z. Cheng, H. Zuo, Z. Ma, X. Peng, X. Chen, Y . Li, E. Cambria, G. Zhao, B. W. Schuller, and J. Tao, “Mer 2025: When affective computing meets large language models,” arXiv preprint arXiv:2504.19423 , 2025

    arXiv 2025

  4. [4]

    Semantically consistent data aug- mentation for neural machine translation via conditional masked language model,

    Q. Cheng, J. Huang, and Y . Duan, “Semantically consistent data aug- mentation for neural machine translation via conditional masked language model,” in Proc. Int. Conf. Comput. Linguist. (COLING) . Int. Comm. Comput. Linguistics, 2022, pp. 5148–5157

    work page 2022

  5. [5]

    Semantics-preserved data aug- mentation for aspect-based sentiment analysis,

    T. Hsu, C. Chen, H. Huang, and H. Chen, “Semantics-preserved data aug- mentation for aspect-based sentiment analysis,” in Proc. Conf. Empirical Methods Nat. Lang. Process. (EMNLP) . Assoc. Comput. Linguistics, 2021, pp. 4417–4422

    work page 2021

  6. [6]

    Refining and synthesis: A simple yet effective data augmentation frame- work for cross-domain aspect-based sentiment analysis,

    H. Wang, K. He, B. Li, L. Chen, F. Li, X. Han, C. Teng, and D. Ji, “Refining and synthesis: A simple yet effective data augmentation frame- work for cross-domain aspect-based sentiment analysis,” in Findings of the Association for Computational Linguistics (ACL) . Assoc. Comput. Linguistics, 2024, pp. 10 318–10 329

    work page 2024

  7. [7]

    DS$^2$-ABSA: Dual-Stream Data Synthesis with Label Refinement for Few-Shot Aspect-Based Sentiment Analysis

    H. Xu, Y . Zhang, Q. Wang, and R. Xu, “Ds 2-absa: Dual-stream data synthesis with label refinement for few-shot aspect-based sentiment analysis,” arXiv preprint arXiv:2412.14849 , 2024

    work page Pith review arXiv 2024

  8. [8]

    Beta distribution guided aspect-aware graph for aspect category sentiment analysis with affective knowledge,

    B. Liang, H. Su, R. Yin, L. Gui, M. Yang, Q. Zhao, X. Yu, and R. Xu, “Beta distribution guided aspect-aware graph for aspect category sentiment analysis with affective knowledge,” in Proc. Conf. Empirical Methods Nat. Lang. Process. (EMNLP) , 2021, pp. 208–218

    work page 2021

Show all 34 references
  1. [9]

    Multi-instance multi-label learning networks for aspect-category sentiment analysis,

    Y . Li, C. Yin, S. Zhong, and X. Pan, “Multi-instance multi-label learning networks for aspect-category sentiment analysis,” in Proc. Conf. Empiri- cal Methods Nat. Lang. Process. (EMNLP) , 2020, pp. 3550–3560

    2020

  2. [10]

    Aspect category sentiment analysis based on prompt-based learning with attention mechanism,

    Z. Ping, G. Sang, Z. Liu, and Y . Zhang, “Aspect category sentiment analysis based on prompt-based learning with attention mechanism,” Neurocomputing, vol. 565, p. 126994, 2024

    2024

  3. [11]

    Solving aspect category sentiment analysis as a text generation task,

    J. Liu, Z. Teng, L. Cui, H. Liu, and Y . Zhang, “Solving aspect category sentiment analysis as a text generation task,” in Proc. Conf. Empirical Methods Nat. Lang. Process. (EMNLP) , 2021, pp. 4406–4416

    2021

  4. [12]

    Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehen- sion,

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer, “Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehen- sion,” in Proc. Annu. Meeting Assoc. Comput. Linguistics (ACL) ...

    2020

  5. [13]

    A comprehensive framework for aspect- category sentiment analysis,

    L. D. Quilio and F. Fioravanti, “A comprehensive framework for aspect- category sentiment analysis,” in Proc. 8th Workshop Natural Lang. Artif. Intell. (NL4AI) at Int. Conf. Italian Assoc. Artif. Intell. (AI*IA) , 2024

    2024

  6. [14]

    Scaling instruction-finetuned language models,

    H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y . Tay, W. Fedus, Y . Li, X. Wang, M. Dehghani, S. Brahma et al., “Scaling instruction-finetuned language models,” J. Mach. Learn. Res. , vol. 25, no. 70, pp. 1–53, 2024

    2024

  7. [15]

    Lego- absa: A prompt-based task assemblable unified generative framework for multi-task aspect-based sentiment analysis,

    T. Gao, J. Fang, H. Liu, Z. Liu, C. Liu, P. Liu, Y . Bao, and W. Yan, “Lego- absa: A prompt-based task assemblable unified generative framework for multi-task aspect-based sentiment analysis,” in Proc. Int. Conf. Comput. Linguistics (COLING), 2022, pp. 7002–7012

    2022

  8. [16]

    Is compound aspect-based sentiment analysis addressed by llms?

    Y . Bai, Z. Han, Y . Zhao, H. Gao, Z. Zhang, X. Wang, and M. Hu, “Is compound aspect-based sentiment analysis addressed by llms?” in Findings Assoc. Comput. Linguistics: EMNLP , 2024, pp. 7836–7861

    2024

  9. [17]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Proc. Conf. North Amer. Chapter Assoc. Comput. Linguistics: Human Lang. Technol. (NAACL-HLT). Assoc. Comput. Linguistics, 2019, pp. 4171– 4186

    2019

  10. [18]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. 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. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B...

    2020

  11. [19]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” in Adv. Neural Inf. Process. Syst., vol. 35, 2022

    2022

  12. [20]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” in Proc. Int. Conf. Learn. Represent. (ICLR). OpenReview.net, 2023

    2023

  13. [21]

    Rvisa: Reasoning and verification for implicit sentiment analysis,

    W. Lai, H. Xie, G. Xu, and Q. Li, “Rvisa: Reasoning and verification for implicit sentiment analysis,” IEEE Trans. Affect. Comput. , 2025

    2025

  14. [22]

    Text data augmentation for large language models: A comprehensive survey of methods, challenges, and opportunities,

    Y . Chai, H. Xie, and S. J. Qin, “Text data augmentation for large language models: A comprehensive survey of methods, challenges, and opportunities,” CoRR, vol. abs/2501.18845, 2025

    2025 arXiv

  15. [23]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” in Proc. Conf. Empirical Methods Nat. Lang. Process. and Int. Joint Conf. Nat. Lang. Process. (EMNLP-IJCNLP) . Assoc. Comput. Linguistics, 2019, pp. 3980–3990

    2019

  16. [24]

    Semeval-2015 task 12: Aspect based sentiment analysis,

    M. Pontiki, D. Galanis, H. Papageorgiou, S. Manandhar, and I. Androut- sopoulos, “Semeval-2015 task 12: Aspect based sentiment analysis,” in Proc. Int. Workshop Semantic Eval. (SemEval@NAACL-HLT) . Assoc. Comput. Linguistics, 2015, pp. 486–495

    2015

  17. [25]

    Semeval-2016 task 5: Aspect based sentiment analysis,

    M. Pontiki, D. Galanis, H. Papageorgiou, I. Androutsopoulos, S. Man- andhar, M. Al-Smadi, M. Al-Ayyoub, Y . Zhao, B. Qin, O. D. Clercq, V . Hoste, M. Apidianaki, X. Tannier, N. V . Loukachevitch, E. V . Kotel- nikov, N. Bel, S. M. Jim ´enez-Zafra, and G. Eryigit, “Semeval-2016...

    2016

  18. [26]

    Enhanced coherence-aware network with hierarchical disentanglement for aspect-category sentiment analysis,

    J. Cui, F. Fukumoto, X. Wang, Y . Suzuki, J. Li, N. Tomuro, and W. Kong, “Enhanced coherence-aware network with hierarchical disentanglement for aspect-category sentiment analysis,” in Proc. Joint Int. Conf. Comput. Linguistics, Lang. Resources and Eval. (LREC-COLING) , 2024, ...

    2024

  19. [27]

    Context-aware embedding for targeted aspect-based sentiment analysis,

    B. Liang, J. Du, R. Xu, B. Li, and H. Huang, “Context-aware embedding for targeted aspect-based sentiment analysis,” in Proc. Annu. Meeting Assoc. Comput. Linguistics (ACL) . Assoc. Comput. Linguistics, 2019, pp. 4678–4683

    2019

  20. [28]

    Sentence constituent-aware aspect-category sentiment analysis with graph attention networks,

    Y . Li, C. Yin, and S. Zhong, “Sentence constituent-aware aspect-category sentiment analysis with graph attention networks,” in Proc. Int. Conf. Natural Lang. Process. and Chinese Comput. (NLPCC) , ser. Lecture Notes in Comput. Sci., vol. 12430. Springer, 2020, pp. 815–827

    2020

  21. [29]

    Locate and combine: A two- stage framework for aspect-category sentiment analysis,

    Y . Wu, Z. Zhang, Y . Zhao, and B. Qin, “Locate and combine: A two- stage framework for aspect-category sentiment analysis,” in Proc. 10th CCF Int. Conf. Natural Lang. Process. Chinese Comput. (NLPCC) , ser. Lect. Notes Comput. Sci., vol. 13028. Springer, 2021, pp. 595–606

    2021

  22. [30]

    Edu-capsule: Aspect-based sentiment analysis at clause level,

    T. Lin, A. Sun, and Y . Wang, “Edu-capsule: Aspect-based sentiment analysis at clause level,” Knowl. Inf. Syst. , vol. 65, no. 2, pp. 517–541, 2023

    2023

  23. [31]

    Aspect-category based sentiment analysis with hierarchical graph convolutional network,

    H. Cai, Y . Tu, X. Zhou, J. Yu, and R. Xia, “Aspect-category based sentiment analysis with hierarchical graph convolutional network,” in Proc. 28th Int. Conf. Comput. Linguistics (COLING) . Int. Comm. Comput. Linguistics, 2020, pp. 833–843

    2020

  24. [32]

    Mvp: Multi-view prompting improves aspect sentiment tuple prediction,

    Z. Gou, Q. Guo, and Y . Yang, “Mvp: Multi-view prompting improves aspect sentiment tuple prediction,” in Proc. 61st Annu. Meeting Assoc. Comput. Linguistics (ACL) , 2023, pp. 4380–4397

    2023

  25. [33]

    Joint aspect and polarity classification for aspect-based sentiment analysis with end-to-end neural networks,

    M. Schmitt, S. Steinheber, K. Schreiber, and B. Roth, “Joint aspect and polarity classification for aspect-based sentiment analysis with end-to-end neural networks,” in Proc. Conf. Empirical Methods Nat. Lang. Process. (EMNLP), 2018, pp. 1109–1114

    2018

  26. [34]

    Dual-attention based joint aspect sentiment clas- sification model,

    P. Gu and Z. Zhang, “Dual-attention based joint aspect sentiment clas- sification model,” in Int. Conf. Web Eng. (ICWE) . Springer, 2022, pp. 252–267

    2022

Pith tools

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