Towards Robust Few-Shot Text Classification Using Transformer Architectures and Dual Loss Strategies

Hongye Zheng; Junliang Du; Weiqiang Huang; Xu Han; Yumeng Sun

REVIEW 4 major objections 5 minor 3 cited by

Towards Robust Few-Shot Text Classification Using Transformer Architectures and Dual Loss Strategies

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding contrastive and L2-regularization losses to cross-entropy improves few-shot text classification, with T5-small reaching 83.7 percent on FewRel 2.0 in the 5-way 5-shot setting.

desk verdict The paper's only ablation for its dual-loss claim is confounded with model and shot count, and the reported numbers lack enough protocol detail to support any of the paper's conclusions. read the letter →

arxiv 2505.06145 v1 pith:WD33P4DT submitted 2025-05-09 cs.CL cs.LG

Xu Han , Yumeng Sun , Weiqiang Huang , Hongye Zheng , Junliang Du This is my paper · ORCID

classification cs.CLcs.LG

keywords few-shottextclassificationFewRel2.0contrastivelearningregularizationTransformerT5-smallrelationcross-entropyloss

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

Few-shot text classification suffers when labeled data is scarce: standard cross-entropy fine-tuning tends to overfit and fails to separate categories with fuzzy boundaries. This paper aims to show that a simple combination of two auxiliary losses—an L2 regularization term and a contrastive term that pulls same-class representations together while pushing different classes apart—added to the usual cross-entropy objective makes Transformer-based few-shot classification more accurate. On FewRel 2.0, the authors report that T5-small reaches 83.7 percent accuracy in the 5-way 5-shot setting, and that the full loss moves 5-way 5-shot accuracy from 72.5 percent with cross-entropy alone to 83.7 percent with both auxiliary terms. The paper also claims that stronger backbones and generative architectures such as T5 benefit most, and that relation categories differ markedly in difficulty, with low-accuracy categories showing the clearest need for contrastive supervision.

What carries the argument

The load-bearing mechanism is the dual-loss training objective $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{CE}} + \lambda \mathcal{L}_{\text{reg}} + \beta \mathcal{L}_{\text{CL}}$. Cross-entropy keeps the model tied to the classification task; the L2 regularization term $\lambda\|\theta\|^2$ is the paper's anti-overfitting device; and the contrastive term, a temperature-scaled softmax over pairwise similarities, reshapes the hidden representation space so same-class samples cluster and different-class samples separate. The paper's reported loss-weight sweep (Figure 3) is the experiment that isolates this mechanism, showing accuracy climbing as the contrastive and regularization weights are added to plain cross-entropy. The Transformer backbones—especially T5-small—supply the feature extractor $h_i = T_\phi(x_i)$ on which those losses act.

What would settle it

Re-run the 5-way 5-shot comparison on FewRel 2.0 using the same pretrained T5-small checkpoint, comparing cross-entropy only against the full loss (CL weight 1, regularization weight 1) over at least 600 randomly sampled episodes and 10 seeds; if the full-loss mean accuracy does not exceed the cross-entropy mean by a margin outside the confidence intervals, the paper's central claim fails.

Watch

Extended reading notes

Core claim

At the center of the paper is the optimization objective $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{CE}} + \lambda \mathcal{L}_{\text{reg}} + \beta \mathcal{L}_{\text{CL}}$, where $\mathcal{L}_{\text{CE}}$ is cross-entropy, $\mathcal{L}_{\text{reg}}$ is an L2 penalty on model parameters, and $\mathcal{L}_{\text{CL}}$ is a supervised contrastive loss with temperature $\tau$. The paper's claim is that this combined objective outperforms plain fine-tuning on FewRel 2.0: in the loss-weight experiment, 5-way 5-shot accuracy rises from 72.5 percent (CE only) to 76.8 percent with contrastive weight 0.5, to 80.3 percent with contrastive weight 1, and to 83.7 percent when regularization at weight 1 is added. A second claim is that the choice of backbone matters: T5-small, DeBERTa-v3, and RoBERTa-base beat BERT-base and ALBERT-base, with T5-small best in every reported setting. A third claim is that categories are not equally hard, with P4 and P8 far below P3 and P6, which the paper interprets as evidence for semantic overlap that cross-entropy alone cannot resolve.

Load-bearing premise

The load-bearing premise is that the accuracy numbers in Table 1 and Figure 3 are stable and representative under FewRel 2.0's episodic evaluation; the paper gives no episode count, random seeds, or variance measure, so the claimed improvement could be an artifact of a single favorable run.

Editorial extensions

If this is right

A practitioner who tunes only the two loss weights can reportedly move FewRel 2.0 5-way 5-shot accuracy from 72.5 percent (cross-entropy only) to 83.7 percent without changing model architecture or data.
T5-small is the recommended backbone among those tested, with the best reported accuracy in every setting, so generative sequence-to-sequence pretraining appears to transfer well to few-shot relation classification.
Adding contrastive loss delivers the largest single jump in the paper's sweep (72.5 to 80.3 percent at weight 1), while regularization adds a further 3.4 points, implying both terms are needed for the best result.
For hard categories like P4 and P8, where accuracy remains below 70 percent, the paper suggests that contrastive and regularization strategies are the right remedy, rather than simply collecting more samples.
In extreme 1-shot and zero-shot conditions, the paper expects generalization challenges to persist, pointing to meta-learning or prompt-based learning as future improvements.

Reading between the lines

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

Editorial inference: because the paper's per-category analysis shows the hardest categories are the ones with overlapping features, a natural testable prediction is that the contrastive term alone should produce the largest accuracy gains on P4 and P8; the paper does not report this per-category ablation.
Editorial inference: if the reported gains replicate across seeds, the same dual-loss recipe is a low-cost drop-in for other low-resource text classification problems; the paper only demonstrates FewRel 2.0.
Editorial inference: the loss sweep suggests a diminishing-returns pattern—contrastive loss gives most of the gain and regularization a smaller tail—so in domains where tuning budget is tight, starting with contrastive loss alone and then adding regularization would be a sensible prior.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a low-evidence empirical note whose central claim is not established. The method is the standard sum of cross-entropy, L2 regularization, and supervised contrastive loss. Nothing here is new, and the one piece of evidence that supposedly shows the loss combination works—Figure 3—is internally inconsistent with Table 1. The CE-only point is 72.5, which matches BERT-base 5-way 1-shot; the final CL:1/Reg:1 point is 83.7, which matches T5-small 5-way 5-shot. The figure never says which model or N-way/K-shot setting was used, so the apparent +11.2 point gain from adding losses could just be the difference between a weaker model in a harder setting and a stronger model in an easier setting. That is a load-bearing confound, not a minor omission.

What the paper does well is modest: the writing is clear, the loss decomposition is easy to follow, and Table 1's model ranking (T5-small > DeBERTa-v3 > RoBERTa-base > BERT-base > ALBERT-base on 5-way) is plausible if the numbers are taken at face value. But the paper gives no episode count, no random seeds, no variance measures, and no evaluation protocol details. The category difficulty analysis in Figure 2 is descriptive and not tied to any hypothesis. The 'adaptive fine-tuning' mentioned in the abstract and method is never defined or tested; it appears to be a label for what is just regular fine-tuning with the stated loss terms.

The citation pattern is also a concern. Several references are to obscure venues or unreported arXiv preprints, and a cluster of them look like self-citations. That by itself would not sink a paper, but it does nothing to shore up the empirical claims.

My verdict is reject. The paper is not incoherent at the paragraph level, but its central assertion rests on an ablation that is confounded and a measurement protocol that is absent. There is no code, no reproducible evidence, and no comparison to existing few-shot baselines. A serious editor should not send this to peer review; it needs a complete rework of the experimental section and a clear statement of what is actually being varied.

For a reading group, not worth the time unless someone wants to use it as an example of how to present ablation results badly. I would not cite it.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a few-shot text classification framework that combines an "adaptive fine-tuning" strategy with contrastive learning and L2 regularization for Transformer models, and evaluates it on the FewRel 2.0 relation-classification benchmark. The experiments compare BERT-base, RoBERTa-base, ALBERT-base, DeBERTa-v3, and T5-small under 5-way/10-way and 1-shot/5-shot settings, report per-category difficulty, and present a loss-weight ablation. The central claim is that the combined objective L_total = CE + lambda*L_reg + beta*L_CL improves few-shot accuracy, with T5-small reaching 83.7% in the 5-way 5-shot setting.

Significance. If the central claim were properly established, the paper would offer a useful practical finding: that adding contrastive and regularization losses to standard cross-entropy fine-tuning improves few-shot relation classification on a widely used benchmark. The paper also usefully draws attention to category-level difficulty variation in FewRel 2.0. However, the manuscript does not provide a defined and tested adaptive fine-tuning method, and the one ablation that supports the loss-synergy claim is confounded with model choice and shot count. As presented, the experimental evidence does not support the stated contribution, so the significance cannot be assessed beyond the level of a plausible hypothesis.

major comments (4)

[Section III-B, Figure 3] The loss-weight sweep that supports the central claim is confounded with model architecture and N-way/K-shot setting. The CE-only endpoint of 72.5% equals the Table 1 accuracy for BERT-base in the 5-way 1-shot setting, while the CL:1/Reg:1 endpoint of 83.7% equals the Table 1 accuracy for T5-small in the 5-way 5-shot setting. The text never states which model or which episodic setting was used for Figure 3, so the reported +11.2-point improvement from adding contrastive and regularization losses cannot be separated from the change in model and shot count. This is the load-bearing evidence for the paper's main claim, and as reported it is internally inconsistent with a controlled ablation.
[Section II, Method] The "adaptive fine-tuning" strategy named in the abstract, introduction, and conclusion is never defined or isolated in the experiments. The equations in Section II describe standard cross-entropy fine-tuning with an L2 penalty on model parameters and a supervised contrastive loss; there is no description of what makes the fine-tuning adaptive, no algorithm, and no comparison that would test this component. The conclusions therefore claim credit for a method component that is neither specified nor measured.
[Section III-B, Table 1 and Figure 3] No experimental protocol is reported for either the model comparison or the loss-weight sweep. The manuscript does not state the number of episodes, random seeds, variance across runs, or whether reported numbers are means over repeated evaluation. Without this information, the single percentages in Table 1 and Figure 3 cannot be assessed for statistical reliability, and known phenomena such as episodic variance in FewRel make the absence of error bars a substantive concern rather than a presentation detail.
[Section III-B, Figure 3 and Section II] The loss weights lambda and beta, as well as the temperature parameter tau, are selected using the reported test accuracies without any validation procedure. The text describes adjusting the weights to obtain better classification results, which makes the reported 83.7% best-case number a form of test-set fitting rather than an evaluation of a pre-specified method. A held-out validation split or a sensitivity analysis over a fixed grid with training-set selection is needed before the synergy claim can be accepted.

minor comments (5)

[Abstract and Keywords] The keyword list contains the typo "Few-short text classification" instead of "Few-shot text classification."
[Section II, optimization objective] The notation is inconsistent: the text says "A is the weight hyperparameter of the contrast loss," but the equation uses beta for that weight. The Greek letters and summation limits in the contrastive-loss formula are also garbled, making the objective difficult to parse.
[Figure 2] The category labels P1 through P8 are not mapped to the corresponding FewRel 2.0 relation names, so the difficulty analysis cannot be checked or reproduced against the dataset.
[References] Reference [19] is listed as "arXiv preprint arXiv: not available," which prevents readers from locating the cited work.
[Section IV, Conclusion] The conclusion states that the findings provide "a theoretical foundation" for future improvements, but the paper contains no theory beyond a standard loss combination; this phrasing overstates what the experiments can support.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper reports direct FewRel 2.0 experiments, and its self-citations are not load-bearing; the ablation confounds are validity concerns, not circularity.

full rationale

The paper contains no derivation chain whose conclusion is equivalent to its inputs. The optimization objective L_total = CE + lambda*L_reg + beta*L_CL is proposed as an empirical recipe, and the claims about it are supported by direct accuracy numbers on the external FewRel 2.0 benchmark (Section III-B, Table 1, Figure 3). No equation is defined in terms of the result it is supposed to establish, and no fitted parameter is relabeled as a prediction. The self-citations (e.g., [16] and [18], which share author J. Du) are used only as background motivation for contrastive learning and hierarchical structure; they do not supply the central empirical claim or forbid alternative methods. The paper's missing protocol details (episode count, seeds, variance) and the apparent alignment of Figure 3 endpoints with different Table 1 rows are serious experimental-validity and reproducibility problems, but they are confounds and omissions, not circular reasoning. Under the rule that circularity must be demonstrated by quote and reduction, no circular step can be identified.

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

The central claim rests on the standard definitions of the three loss terms, the implicit assumption that the FewRel benchmark is used correctly (random episode sampling, enough episodes for reliable estimates), and a set of unspecified hyperparameters (lambda, beta, tau, plus training details) whose values are not disclosed. No code or data is shipped, and the best loss-weight combination is chosen from the reported accuracies, which makes the conclusion partly fit to the test set.

free parameters (4)

lambda (weight decay coefficient)
Appears in L_reg = lambda * ||theta||^2 but its value is never given.
beta (contrast loss weight) = varied: 0, 0.5, 1 in Figure 3
The paper selects the best value from test accuracy, making the reported gain dependent on this fitted choice.
tau (temperature in contrastive loss)
Not specified; the contrastive loss formula includes it.
Training hyperparameters (learning rate, batch size, number of episodes)
None reported; required to reproduce any accuracy number.

assumptions (4)

standard math The cross-entropy, contrastive, and L2 losses are implemented in their standard forms and combined linearly.
The paper defines these losses in Section II and uses them without modification.
domain assumption The FewRel 2.0 dataset is evaluated under the standard N-way K-shot episodic protocol with random category and example sampling.
Section III-A says they follow the predefined framework, but no episode count or sampling details are given.
domain assumption The reported accuracies are stable estimates despite the absence of error bars.
All conclusions in Section III-B compare single accuracy numbers; without variance, this assumption is load-bearing.
domain assumption The pre-trained models are used as distributed and fine-tuned without additional architectural changes.
The method section mentions LoRA and hierarchical terms, but the experiments do not implement them, so the actual system is just standard fine-tuning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Robust Few-Shot Text Classification Using Transformer Architectures and Dual Loss Strategies." pith.science (2026). https://pith.science/paper/WD33P4DT

@misc{pith2026250506145,
  author       = {Pith},
  title        = {Pith review of: Towards Robust Few-Shot Text Classification Using Transformer Architectures and Dual Loss Strategies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WD33P4DT}},
  note         = {Machine review of arXiv:2505.06145}
}

read the original abstract

Few-shot text classification has important application value in low-resource environments. This paper proposes a strategy that combines adaptive fine-tuning, contrastive learning, and regularization optimization to improve the classification performance of Transformer-based models. Experiments on the FewRel 2.0 dataset show that T5-small, DeBERTa-v3, and RoBERTa-base perform well in few-shot tasks, especially in the 5-shot setting, which can more effectively capture text features and improve classification accuracy. The experiment also found that there are significant differences in the classification difficulty of different relationship categories. Some categories have fuzzy semantic boundaries or complex feature distributions, making it difficult for the standard cross entropy loss to learn the discriminative information required to distinguish categories. By introducing contrastive loss and regularization loss, the generalization ability of the model is enhanced, effectively alleviating the overfitting problem in few-shot environments. In addition, the research results show that the use of Transformer models or generative architectures with stronger self-attention mechanisms can help improve the stability and accuracy of few-shot classification.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Time-Series Learning for Proactive Fault Prediction in Distributed Systems with Deep Neural Structures
cs.DC 2025-05 conditional novelty 3.0 of 10

A GRU plus attention plus feedforward classifier outperforms transformer baselines on Azure telemetry fault prediction in the reported metrics, but without code, error bars, or train/test details.
Capsule Network-Based Semantic Intent Modeling for Human-Computer Interaction
cs.CL 2025-07 reject novelty 2.0 of 10

A capsule network with dynamic routing is reported to classify SNIPS intents with 95.6% accuracy, but uncontrolled baselines and missing experimental details undermine the claim.
Structured Memory Mechanisms for Stable Context Representation in Large Language Models
cs.CL 2025-05 reject novelty 2.0 of 10

A gated memory module with attention-based reading and forgetting is reported to improve NarrativeQA and dialogue consistency over GPT-2, BART, Longformer, and RETRO.

Reference graph

Works this paper leans on

25 extracted references · 9 canonical work pages · cited by 3 Pith papers

[1]

Transformers are short-text classifiers,

F. Karl and A. Scherp, “Transformers are short-text classifiers,” Proceedings of the International Cross-Domain Conference for Machine Learning and Knowledge Extraction, Cham: Springer Nature Switzerland, pp. 103–122, 2023

work page 2023
[2]

Medical Entity-Driven Analysis of Insurance Claims Using a Multimodal Transformer Model,

X. Wang, “Medical Entity-Driven Analysis of Insurance Claims Using a Multimodal Transformer Model,” Journal of Computer Technology and Software, vol. 4, no. 3, 2025

2025
[3]

A Self-Supervised Vision Transformer Approach for Dermatological Image Analysis,

F. Guo, X. Wu, L. Zhang, H. Liu and A. Kai, “A Self-Supervised Vision Transformer Approach for Dermatological Image Analysis,” Journal of Computer Science and Software Applications, vol. 5, no. 4, 2025

2025
[4]

A Visual Communication Optimization Method for Human- Computer Interaction Interfaces Using Fuzzy Logic and Wavelet Transform,

Q. Sun, “A Visual Communication Optimization Method for Human- Computer Interaction Interfaces Using Fuzzy Logic and Wavelet Transform,” Proceedings of the 2024 4th International Conference on Communication Technology and Information Technology (ICCTIT), pp. 140–144, Dec. 2024

2024
[5]

Social Network User Profiling for Anomaly Detection Based on Graph Neural Networks,

Y. Zhang, “Social Network User Profiling for Anomaly Detection Based on Graph Neural Networks,” arXiv preprint arXiv:2503.19380, 2025

arXiv 2025
[6]

Unsupervised Detection of Fraudulent Transactions in E-commerce Using Contrastive Learning,

X. Li, Y. Peng, X. Sun, Y. Duan, Z. Fang and T. Tang, “Unsupervised Detection of Fraudulent Transactions in E-commerce Using Contrastive Learning,” arXiv preprint arXiv:2503.18841, 2025

arXiv 2025
[7]

A survey of text classification with transformers: How wide? how large? how long? how accurate? how expensive? how safe?

J. Fields, K. Chovanec and P. Madiraju, “A survey of text classification with transformers: How wide? how large? how long? how accurate? how expensive? how safe?” IEEE Access, vol. 12, pp. 6518–6531, 2024

work page 2024
[8]

Transformer-Based Structural Anomaly Detection for Video File Integrity Assessment,

D. Xu, “Transformer-Based Structural Anomaly Detection for Video File Integrity Assessment,” Transactions on Computational and Scientific Methods, vol. 5, no. 4, 2024

2024

Show all 25 references

[9]

A Graph Attention-Based Recommendation Framework for Sparse User-Item Interactions,

A. Liang, “A Graph Attention-Based Recommendation Framework for Sparse User-Item Interactions,” Journal of Computer Science and Software Applications, vol. 5, no. 4, 2025

2025
[10]

Addressing Class Imbalance with Probabilistic Graphical Models and Variational Inference,

Y. Lou, J. Liu, Y. Sheng, J. Wang, Y. Zhang and Y. Ren, “Addressing Class Imbalance with Probabilistic Graphical Models and Variational Inference,” arXiv preprint arXiv:2504.05758, 2025

2025 arXiv
[11]

Single-Device Human Activity Recognition Based on Spatiotemporal Feature Learning Networks,

J. Zhan, “Single-Device Human Activity Recognition Based on Spatiotemporal Feature Learning Networks,” Transactions on Computational and Scientific Methods, vol. 5, no. 3, 2025

2025
[12]

Human-Computer Interaction in Smart Devices: Leveraging Sentiment Analysis and Knowledge Graphs for Personalized User Experiences,

S. Duan, “Human-Computer Interaction in Smart Devices: Leveraging Sentiment Analysis and Knowledge Graphs for Personalized User Experiences,” Proceedings of the 2024 4th International Conference on Electronic Information Engineering and Computer Communication (EIECC), pp. 1294...

2024
[13]

Distributed Network Traffic Scheduling via Trust-Constrained Policy Learning Mechanisms,

Y. Ren, M. Wei, H. Xin, T. Yang and Y. Qi, “Distributed Network Traffic Scheduling via Trust-Constrained Policy Learning Mechanisms,” Transactions on Computational and Scientific Methods, vol. 5, no. 4, 2024

2024
[14]

A Reinforcement Learning Approach to Traffic Scheduling in Complex Data Center Topologies,

Y. Deng, “A Reinforcement Learning Approach to Traffic Scheduling in Complex Data Center Topologies,” Journal of Computer Technology and Software, vol. 4, no. 3, 2025

2025
[15]

Optimizing Distributed Computing Resources with Federated Learning: Task Scheduling and Communication Efficiency,

Y. Wang, “Optimizing Distributed Computing Resources with Federated Learning: Task Scheduling and Communication Efficiency,” Journal of Computer Technology and Software, vol. 4, no. 3, 2025

2025
[16]

Contrastive and Variational Approaches in Self-Supervised Learning for Complex Data Mining,

Y. Liang, L. Dai, S. Shi, M. Dai, J. Du and H. Wang, “Contrastive and Variational Approaches in Self-Supervised Learning for Complex Data Mining,” arXiv preprint arXiv:2504.04032, 2025

2025 arXiv
[17]

Efficient Compression of Large Language Models with Distillation and Fine-Tuning,

A. Kai, L. Zhu and J. Gong, “Efficient Compression of Large Language Models with Distillation and Fine-Tuning,” Journal of Computer Science and Software Applications, vol. 3, no. 4, pp. 30–38, 2023

2023
[18]

Investigating Hierarchical Term Relationships in Large Language Models,

G. Cai, J. Gong, J. Du, H. Liu and A. Kai, “Investigating Hierarchical Term Relationships in Large Language Models,” Journal of Computer Science and Software Applications, vol. 5, no. 4, 2025

2025
[19]

Revisiting LoRA: A Smarter Low-Rank Approach for Efficient Model Adaptation,

Y. Wang, Z. Fang, Y. Deng, L. Zhu, S. Duan and Y. Peng, “Revisiting LoRA: A Smarter Low-Rank Approach for Efficient Model Adaptation,” arXiv preprint arXiv: not available, 2025

2025
[20]

A comparative survey of instance selection methods applied to non-neural and transformer-based text classification,

W. Cunha, F. Viegas, C. França, et al., “A comparative survey of instance selection methods applied to non-neural and transformer-based text classification,” ACM Computing Surveys, vol. 55, no. 13s, pp. 1–52, 2023

2023
[21]

Deep Learning for Cross-Domain Recommendation with Spatial-Channel Attention,

L. Zhu, “Deep Learning for Cross-Domain Recommendation with Spatial-Channel Attention,” Journal of Computer Science and Software Applications, vol. 5, no. 4, 2025

2025
[22]

Mask-guided BERT for few-shot text classification,

W. Liao, Z. Liu, H. Dai, et al., “Mask-guided BERT for few-shot text classification,” Neurocomputing, vol. 610, pp. 128576, 2024

2024
[23]

Short‐ Text Classification Detector: A BERT ‐ Based Mental Approach,

Y. Hu, J. Ding, Z. Dou, et al., “ Short‐ Text Classification Detector: A BERT ‐ Based Mental Approach, ” Computational Intelligence and Neuroscience, vol. 2022, no. 1, pp. 8660828, 2022

2022
[24]

Fine-Tuned ‘Small’ LLMs (Still) Significantly Outperform Zero-Shot Generative AI Models in Text Classification,

M. J. J. Bucher and M. Martini, “Fine-Tuned ‘Small’ LLMs (Still) Significantly Outperform Zero-Shot Generative AI Models in Text Classification,” arXiv preprint arXiv:2406.08660, 2024

2024 arXiv
[25]

Continual few-shot learning for text classification,

R. Pasunuru, V. Stoyanov and M. Bansal, “Continual few-shot learning for text classification,” Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 5688–5702, 2021

2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Transformers are short-text classifiers,

F. Karl and A. Scherp, “Transformers are short-text classifiers,” Proceedings of the International Cross-Domain Conference for Machine Learning and Knowledge Extraction, Cham: Springer Nature Switzerland, pp. 103–122, 2023

work page 2023

[2] [2]

Medical Entity-Driven Analysis of Insurance Claims Using a Multimodal Transformer Model,

X. Wang, “Medical Entity-Driven Analysis of Insurance Claims Using a Multimodal Transformer Model,” Journal of Computer Technology and Software, vol. 4, no. 3, 2025

2025

[3] [3]

A Self-Supervised Vision Transformer Approach for Dermatological Image Analysis,

F. Guo, X. Wu, L. Zhang, H. Liu and A. Kai, “A Self-Supervised Vision Transformer Approach for Dermatological Image Analysis,” Journal of Computer Science and Software Applications, vol. 5, no. 4, 2025

2025

[4] [4]

A Visual Communication Optimization Method for Human- Computer Interaction Interfaces Using Fuzzy Logic and Wavelet Transform,

Q. Sun, “A Visual Communication Optimization Method for Human- Computer Interaction Interfaces Using Fuzzy Logic and Wavelet Transform,” Proceedings of the 2024 4th International Conference on Communication Technology and Information Technology (ICCTIT), pp. 140–144, Dec. 2024

2024

[5] [5]

Social Network User Profiling for Anomaly Detection Based on Graph Neural Networks,

Y. Zhang, “Social Network User Profiling for Anomaly Detection Based on Graph Neural Networks,” arXiv preprint arXiv:2503.19380, 2025

arXiv 2025

[6] [6]

Unsupervised Detection of Fraudulent Transactions in E-commerce Using Contrastive Learning,

X. Li, Y. Peng, X. Sun, Y. Duan, Z. Fang and T. Tang, “Unsupervised Detection of Fraudulent Transactions in E-commerce Using Contrastive Learning,” arXiv preprint arXiv:2503.18841, 2025

arXiv 2025

[7] [7]

A survey of text classification with transformers: How wide? how large? how long? how accurate? how expensive? how safe?

J. Fields, K. Chovanec and P. Madiraju, “A survey of text classification with transformers: How wide? how large? how long? how accurate? how expensive? how safe?” IEEE Access, vol. 12, pp. 6518–6531, 2024

work page 2024

[8] [8]

Transformer-Based Structural Anomaly Detection for Video File Integrity Assessment,

D. Xu, “Transformer-Based Structural Anomaly Detection for Video File Integrity Assessment,” Transactions on Computational and Scientific Methods, vol. 5, no. 4, 2024

2024