Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Lowering the Barrier of Machine Learning: Achieving Zero Manual Labeling in Review Classification Using LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Even with zero manually labeled data, an LLM can pick and label a small training set that pushes classical classifiers to 80.8–88.6% accuracy on review sentiment.

desk verdict Plausible zero-labeling pipeline, but an undefined data split and likely GPT-4 benchmark contamination leave the central accuracy claims unsupported. read the letter →

arxiv 2502.02893 v1 pith:EIDPKPMR submitted 2025-02-05 cs.CL

classification cs.CL
keywords sentimentclassificationlargelanguagemodelszero-shotlabelingtrainingsetgenerationreviewanalysisGPT-4BERTdomainadaptationlow-resourceNLP
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 claims that the manual annotation bottleneck in review sentiment classification can be removed entirely. A prompt-configured LLM, ESCS-GPT, reads unlabeled reviews, picks the 100 examples it judges most useful for training, and labels them; domain-adapted BERT-based embedders (URSLM) convert reviews into vectors; and ordinary classifiers — SVM, decision tree, random forest, and logistic regression — do the final classification. Across the Movie, TripAdvisor, and Amazon review datasets, the best pipeline reaches 83.4, 88.6, and 80.8 percent accuracy respectively in five-fold cross-validation, without any manually labeled training data. The paper reports that this outperforms classic TF-IDF and bag-of-words baselines trained on 100 randomly sampled human-labeled reviews and closely approaches or matches LLM-based baselines trained on the original labels. If accurate, the result matters because it would let non-experts with modest hardware build useful sentiment classifiers through a chat conversation.

What carries the argument

The machinery is a pipeline with three components. ESCS-GPT is a prompt-configured GPT-4 assistant given role instructions and a task description that asks it to select and label the most informative reviews; it produces a 100-instance labeled training set per fold. URSLM is a RoBERTa-base or ALBERT-baseV2 model further pre-trained on a domain corpus of 10,000 reviews per dataset using masked language modeling, then wrapped with a mean-pooling layer to turn each review into a fixed-length embedding. The final stage uses standard classifiers — SVM with an RBF kernel, decision tree, random forest, and logistic regression — trained on the URSLM embeddings. The argument's load is carried by the claim that LLM-selected and LLM-labeled examples teach simple classifiers at least as well as randomly selected human-labeled examples do.

What would settle it

Run the same pipeline on a freshly written set of reviews that postdates GPT-4's training cutoff and whose labels have never been public; if accuracy falls to the level of the random-labeled baselines, the result is explained by memorization rather than label-free selection. Separately, check for exact-string overlap between the 100 selected training reviews in each fold and that fold's test reviews; any overlap invalidates the reported test accuracy.

Watch

Extended reading notes

Core claim

The central claim is that zero-manual-labeling sentiment classification is achievable by using an LLM as a training-set generator rather than as the final classifier. ESCS-GPT, configured through prompts, is given unlabeled reviews and asked to select and label the 100 most valuable instances for a downstream classifier. Those instances are embedded with URSLM, a RoBERTa-base or ALBERT-baseV2 model further pre-trained on review text, and then fed to one of four classical classifiers. On the Movie, TripAdvisor, and Amazon datasets, the best pipeline configuration achieves average accuracies of 83.4, 88.6, and 80.8 percent, respectively, and the paper states that the approach significantly outperformed the classic baselines on the Movie and Amazon datasets and performed better than the classic baselines on the TripAdvisor dataset, while matching or closely approximating LLM-based baselines that used the original labels. The implication the author draws is that manual labeling, machine-learning expertise, and large compute budgets are not prerequisites for high-accuracy review classification.

Load-bearing premise

The argument breaks down if the language model already knows the correct labels for these public benchmark reviews, or if any review it selects for training is also used for testing.

Editorial extensions

If this is right

  • A small business could obtain a review classifier by chatting with an LLM to build a 100-example labeled set and then training a logistic regression or random forest on embeddings, all within free cloud-notebook memory limits.
  • The approach removes the per-domain annotation cost: the same prompt and pipeline are applied to retail, service, and cultural review domains with only a dataset change, and no manual labels are used in any of them.
  • Classic feature methods (TF-IDF and bag-of-words) fall behind on Movie and Amazon, so the gains come from the LLM's selection and labeling rather than from the embeddings alone.
  • LLM-based baselines that use the original human labels reach only slightly higher accuracy on these datasets, so the practical gap between zero-label and fully labeled pipelines is small.

Reading between the lines

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

  • Because GPT-4's training data almost certainly includes these widely used public benchmark reviews, the zero-manual-labeling result may not transfer to fresh, private reviews; a true test requires reviews written after the model's knowledge cutoff.
  • A concrete extension would be to run the same ESCS-GPT prompt on an open-weight LLM and on a private, never-published review set, and to compare accuracy with and without exact-match leakage checks.
  • The same generate-a-small-training-set pattern could be reused for other downstream classification tasks where an LLM can judge label quality, such as spam detection or support-ticket routing, with the same caveat about benchmark contamination.
  • Because the pipeline relies on a proprietary chat API, reproducibility and cost depend on external service terms; a reproducible variant would need to log every prompt, selection, and model version.
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 / 5 minor

Summary. The paper proposes a sentiment-classification pipeline that removes manual labeling: ESCS-GPT (a ChatGPT/GPT-4 configuration) selects and labels 100 customer reviews per cross-validation fold, URSLM (further pre-trained ALBERT/RoBERTa) converts reviews into fixed-length embeddings, and standard classifiers (SVM, DT, RF, LR) perform the final classification. The authors evaluate the pipeline on three public review datasets (Movie, TripAdvisor, Amazon) with 5-fold cross-validation and report accuracies of 80.8–88.6% for the best pipeline per dataset. They compare against classical feature-based baselines (TF-IDF/BoW) and LLM-based baselines (base RoBERTa/ALBERT) that use 100 randomly sampled human-labeled instances per fold, concluding that the approach outperforms classical baselines without any manually labeled data while remaining computationally light enough to run on free Colab resources.

Significance. If the central claim holds, this would be a practically useful contribution: it would let non-experts build competitive review classifiers without annotation effort, using only prompt-based LLM interaction and modest compute. The paper has clear strengths: the accessibility goal is well motivated; the evaluation uses three domains with different label distributions; and RQ2 reports concrete memory and time measurements, giving readers a realistic picture of resource needs. The modular design also makes the approach easy to describe. However, the evaluation as written does not yet establish the central claim. The undefined "Process Data Set" raises a possible train/test leakage path, the closed LLM may be contaminated by the public benchmark data, and the RQ3 comparisons conflate label source, instance selection, and embedding adaptation. These are load-bearing issues that require either clarification with additional experiments or a substantial narrowing of the claims.

major comments (4)
  1. [V.C, V.E] The term "Process Data Set" is used in Section V.C and Section V.E but is never defined. The query protocol states that "the instances of the Process Data Set were provided as query attachments to ESCS-GPT after their labels were removed," and the baseline protocol in Section V.E states that "we randomly sampled 100 instances from the Process Data Set." If this set is the entire 5,000-instance Experimental Dataset rather than the training portion of the current fold, then in every fold the ESCS-GPT-selected instances and the baseline random samples can include reviews from the held-out test fold, and the classifiers are trained and evaluated on overlapping instances. That would inflate every accuracy value in Table VI and invalidate all comparisons in Figure 6 and Table VIII. The paper must define this set explicitly, state that each fold uses only the training fold for instance selection and labeling, and ideally release the split and selection code so the protocol can be verified.
  2. [VI] Section VI's discussion of pre-training bias does not address benchmark contamination. The Amazon, Movie (IMDB), and TripAdvisor datasets are widely used public benchmarks, so GPT-4 may have encountered them or near-duplicates during pre-training. If so, ESCS-GPT's labels could agree with the ground-truth labels by memorization rather than by a genuinely label-free process, making the reported accuracies an artifact of benchmark overlap rather than evidence for the zero-labeling claim. The paper should either present a concrete test on private or newly collected review data, compare ESCS-GPT's label accuracy on subsets with high versus low likely overlap with GPT-4 training data, or otherwise provide evidence that the generated labels are not informed by the evaluation ground truth.
  3. [V.E] The RQ3 comparison confounds three factors that are each sufficient to explain performance differences: label source (ESCS-GPT labels versus original human labels), instance selection (ESCS-GPT's "most valuable" selection versus random 100-instance sampling), and embedding adaptation (URSLM further pre-training versus base RoBERTa/ALBERT). The baselines in Table VIII use random sampling and base models, so the observed gains over classic baselines cannot be attributed to the zero-labeling mechanism alone. The paper needs ablations that vary one factor at a time (for example, ESCS-GPT labels with random selection, human labels with ESCS-GPT selection, and URSLM embeddings with human labels) to support the conclusion that eliminating manual labeling is what preserves accuracy.
  4. [V.E] The text states that the approach "significantly outperformed" the classic baseline models on the Movie and Amazon datasets, but no statistical significance test is reported. The standard deviations in Table VI are over five folds, and the baseline standard deviations in Table VIII are over random samplings; they do not directly support a paired claim about the difference between pipelines. A paired test over folds (or a confidence interval for the per-fold accuracy difference) is needed to justify the word "significantly."
minor comments (5)
  1. [Table II] The prompt text "downstream segment classifiers" appears to be a typo for "downstream sentiment classifiers."
  2. [Section V.C] The ESCS-GPT query in Table V asks ChatGPT to "provide download links" for the selected reviews; this is an implementation detail that should be replaced with a precise, reproducible specification of what the model returns and how that output is parsed into the training set.
  3. [Figure 6] Figure 6 shows only point values for the average accuracies; adding error bars or confidence intervals would be helpful given that Table VIII reports substantial standard deviations for several baselines.
  4. [References] Reference [22] contains the formatting error "Sastry, A. Askellet al."; the author name should be "Askell et al." with proper spacing.
  5. [VI] The internal-validity discussion states that there is no overlap between the BERT-model training data and the experimental data, but the paper does not explain how the "not chosen as the Domain Corpus" sampling in Section IV.A was performed. A brief description of the sampling procedure would make the no-overlap claim easier to verify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: reported accuracies are empirical outcomes of an LLM-labeling pipeline tested against independent human labels.

full rationale

The paper's derivation chain is not circular. ESCS-GPT selects and labels 100 reviews from the data provided to it; URSLMs embed both the resulting training set and the test set; classical classifiers are trained on the LLM-produced labels; and accuracy/F1/recall are computed against the original human-provided labels in the Experimental Dataset. At no point does the paper fit a parameter to a subset of the evaluation labels and then 'predict' that same subset: the LLM-generated labels are inputs to the downstream classifiers, while the evaluation labels are independent outputs used only for scoring. No equation in the paper defines a predicted quantity in terms of its ground truth, and no fitted value is renamed as a prediction. The paper also does not rest on self-citations or on any uniqueness theorem imported from the authors' prior work; all cited support is external and none of it is load-bearing for the central empirical comparison. The undefined 'Process Data Set' in Section V.C and the possibility that GPT-4 was pretrained on the public benchmark datasets are genuine threats to internal and external validity, and they should be addressed with a precise split specification and contamination analysis. However, per the hard rules, those concerns are correctness risks rather than circularity, because the paper's text does not exhibit a construction-level equivalence between an input and an output. The reported accuracies (80.8-88.6%) are empirical measurements, not consequences of the method's definitions, so the honest finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on three assumptions that the paper does not test: that GPT-4's labels are independent of the benchmark ground truth, that ESCS-GPT sees only training-fold instances, and that the original dataset labels are correct. The free parameters are the hand-chosen training-set size, prompt design, and pre-training hyperparameters. No new entities are invented; ESCS-GPT and URSLM are named configurations of existing models (GPT-4 and RoBERTa or ALBERT).

free parameters (5)
  • Number of LLM-selected training instances per fold = 100
    Chosen by the authors without justification (Section V.C). All reported accuracies depend on a training set of exactly 100 instances.
  • MLM further pre-training epochs = 3
    Set in Table III and used for both RoBERTa and ALBERT; no sensitivity analysis is reported.
  • MLM batch size = 14
    Listed in Table III; chosen without stated justification.
  • Decision tree max depth = 10
    Set in Section IV.B.3 to prevent overfitting; not varied.
  • ESCS-GPT prompt wording = Multiple iterations of prompt engineering
    The prompts in Tables II and V were refined through iterative testing (Section IV.B.1), making the prompt design a hand-tuned component of the method.
assumptions (4)
  • domain assumption The original labels in the Movie, TripAdvisor, and Amazon datasets are accurate ground truth for sentiment.
    Used to measure accuracy in Section V and to generate baseline labels in RQ3; mislabeled ground truth would bias all comparisons.
  • ad hoc to paper ESCS-GPT's generated labels are independent of the ground-truth labels used for evaluation.
    The paper assumes GPT-4 does not use memorized test-set knowledge to label the reviews; this is not verified (Sections V and VI).
  • ad hoc to paper The Process Data Set supplied to ESCS-GPT contains only the training fold of the current cross-validation split.
    Section V.C describes providing the instances of the Process Data Set without defining this set or stating that test folds are withheld, introducing a potential leakage path.
  • standard math Standard supervised machine learning background: classifiers trained on labeled vectors generalize to held-out instances from the same distribution.
    Implicitly assumed in the cross-validation evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lowering the Barrier of Machine Learning: Achieving Zero Manual Labeling in Review Classification Using LLMs." pith.science (2026). https://pith.science/paper/EIDPKPMR

@misc{pith2026250202893,
  author       = {Pith},
  title        = {Pith review of: Lowering the Barrier of Machine Learning: Achieving Zero Manual Labeling in Review Classification Using LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EIDPKPMR}},
  note         = {Machine review of arXiv:2502.02893}
}
read the original abstract

With the internet's evolution, consumers increasingly rely on online reviews for service or product choices, necessitating that businesses analyze extensive customer feedback to enhance their offerings. While machine learning-based sentiment classification shows promise in this realm, its technical complexity often bars small businesses and individuals from leveraging such advancements, which may end up making the competitive gap between small and large businesses even bigger in terms of improving customer satisfaction. This paper introduces an approach that integrates large language models (LLMs), specifically Generative Pre-trained Transformer (GPT) and Bidirectional Encoder Representations from Transformers (BERT)-based models, making it accessible to a wider audience. Our experiments across various datasets confirm that our approach retains high classification accuracy without the need for manual labeling, expert knowledge in tuning and data annotation, or substantial computational power. By significantly lowering the barriers to applying sentiment classification techniques, our methodology enhances competitiveness and paves the way for making machine learning technology accessible to a broader audience.

Figures

Figures reproduced from arXiv: 2502.02893 by the authors.

Figure 1
Figure 1. Overall workflow of our approach classification in downstream tasks. Finally, we integrated mul￾tiple machine learning classifiers to perform the sentiment classification tasks, further ensuring the robustness of our approach across different application scenarios. In conclusion, our approach is designed to eliminate the need for manual labeling, expert knowledge in tuning and data annotation, and to reduce the reli… view at source ↗
Figure 2
Figure 2. Workflow of data preparation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Length distribution of Domain Corpus task by excluding reviews with a neutral rating (3) and labeling reviews with poor ratings (1-2) as negative sentiment (’0’) and those with good ratings (4-5) as positive sentiment (’1’). The length distributions of each dataset within the Exper￾imental Dataset are illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Length distribution across datasets in Table II. The design of our ESCS-GPT configuration com￾mands (prompts), based on established principles in prompt engineering, including role assignment and detailed task de￾scriptions, was refined through multiple iterations to e…
Figure 5
Figure 5. Figure 5: Training loss [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: RQ3 comparative experiment results reviews dataset from Stanford University, specifically designed for sentiment classification. We also addressed configuration command issues by referring to relevant prompt engineering studies and conducting iterative testing to ensur…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    The effect of word of mouth on sales: Online book reviews,

    J. A. Chevalier and D. Mayzlin, “The effect of word of mouth on sales: Online book reviews,” Journal of Marketing Research , vol. 43, no. 3, pp. 345–354, 2006

  2. [2]

    Word of mouth for movies: Its dynamics and impact on box office revenue,

    Y . Liu, “Word of mouth for movies: Its dynamics and impact on box office revenue,” Journal of Marketing , vol. 70, no. 3, pp. 74–89, 2006

  3. [3]

    Overcoming the j-shaped distribution of product reviews,

    N. Hu, P. A. Pavlou, and J. Zhang, “Overcoming the j-shaped distribution of product reviews,” Communications of the ACM , vol. 52, no. 10, pp. 144–147, 2009

  4. [4]

    The digitization of word of mouth: Promise and chal- lenges of online feedback mechanisms,

    C. Dellarocas, “The digitization of word of mouth: Promise and chal- lenges of online feedback mechanisms,” Management Science, vol. 49, no. 10, pp. 1407–1424, 2003

  5. [5]

    Examining the relationship between reviews and sales: The role of reviewer identity disclosure in electronic markets,

    C. Forman, A. Ghose, and B. Wiesenfeld, “Examining the relationship between reviews and sales: The role of reviewer identity disclosure in electronic markets,” Information Systems Research , vol. 19, no. 3, pp. 291–313, 2008

  6. [6]

    An analysis of sentiment: Methods, applications, and challenges,

    H. D. Sharma and P. Goyal, “An analysis of sentiment: Methods, applications, and challenges,” Engineering Proceedings, vol. 59, no. 1, p. 68, 2023

  7. [7]

    Sentiment analysis: An overview from linguistics,

    M. Taboada, “Sentiment analysis: An overview from linguistics,” Annual Review of Linguistics , vol. 2, 02 2016

  8. [8]

    Sentiment analysis: A survey on design framework, applications and future scopes,

    M. Bordoloi and S. Biswas, “Sentiment analysis: A survey on design framework, applications and future scopes,” Artificial Intelligence Re- view, vol. 56, 03 2023

Show all 37 references
  1. [9]

    Machine learning algorithms for labeling: Where and how they are used?

    T. Fredriksson, J. Bosch, H. H. Olsson, and D. I. Mattos, “Machine learning algorithms for labeling: Where and how they are used?” in 2022 IEEE International Systems Conference (SysCon) , 2022, pp. 1–8

  2. [10]

    Cheap and fast— but is it good? evaluating non-expert annotations for natural language tasks,

    R. Snow, B. O’Connor, D. Jurafsky, and A. Y . Ng, “Cheap and fast— but is it good? evaluating non-expert annotations for natural language tasks,” Proceedings of the conference on empirical methods in natural language processing (EMNLP) , pp. 254–263, 2008

  3. [11]

    A review on large language models: Architectures, applications, taxonomies, open issues and challenges,

    M. A. K. Raiaan, M. S. H. Mukta, K. Fatema, N. M. Fahad, S. Sakib, M. M. J. Mim, J. Ahmad, M. E. Ali, and S. Azam, “A review on large language models: Architectures, applications, taxonomies, open issues and challenges,” IEEE Access, vol. 12, pp. 26 839–26 874, 2024

  4. [12]

    Research and implementation of high computational power for training and inference of convolutional neural networks,

    T. Li, B. He, and Y . Zheng, “Research and implementation of high computational power for training and inference of convolutional neural networks,” Applied Sciences , vol. 13, no. 2, 2023. [Online]. Available: https://www.mdpi.com/2076-3417/13/2/1003

  5. [13]

    Lexicon- based methods for sentiment analysis,

    M. Taboada, J. Brooke, M. Tofiloski, K. V oll, and M. Stede, “Lexicon- based methods for sentiment analysis,” Computational linguistics , vol. 37, no. 2, pp. 267–307, 2011

  6. [14]

    Thumbs up?: sentiment classifi- cation using machine learning techniques,

    B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up?: sentiment classifi- cation using machine learning techniques,” in Proceedings of the ACL- 02 conference on Empirical methods in natural language processing- Volume 10. Association for Computational Linguistics, 2002, pp. 79– 86

  7. [15]

    Machine learning in automated text categorization,

    F. Sebastiani, “Machine learning in automated text categorization,” ACM computing surveys (CSUR) , vol. 34, no. 1, pp. 1–47, 2002

  8. [16]

    Convolutional neural networks for sentence classification,

    Y . Kim, “Convolutional neural networks for sentence classification,” in Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , 2014, pp. 1746–1751

  9. [17]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

  10. [18]

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

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

  11. [19]

    Deep learning for sentiment analysis: A survey,

    L. Zhang, S. Wang, and B. Liu, “Deep learning for sentiment analysis: A survey,” Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 8, no. 4, p. e1253, 2018

  12. [20]

    Xlnet: Generalized autoregressive pretraining for language understanding,

    Z. Yang, Z. Dai, Y . Yang, J. Carbonell, R. R. Salakhutdinov, and Q. V . Le, “Xlnet: Generalized autoregressive pretraining for language understanding,” arXiv preprint arXiv:1906.08237 , 2019

  13. [21]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” 2009 IEEE conference on computer vision and pattern recognition , pp. 248–255, 2009

  14. [22]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” arXiv preprint arXiv:2005.14165 , 2020

  15. [23]

    Energy and policy considerations for deep learning in NLP,

    E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in NLP,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , A. Korhonen, D. Traum, and L. M `arquez, Eds. Florence, Italy: Association for Com...

  16. [24]

    Frugalgpt: How to use large language models while reducing cost and improving performance,

    L. Chen, M. Zaharia, and J. Zou, “Frugalgpt: How to use large language models while reducing cost and improving performance,”

  17. [25]

    Memory-efficient pipeline-parallel dnn training,

    D. Narayanan, A. Phanishayee, K. Shi, X. Chen, and M. Zaharia, “Memory-efficient pipeline-parallel dnn training,” vol. 139, pp. 7937– 7947, 18–24 Jul 2021. [Online]. Available: https://proceedings.mlr. press/v139/narayanan21a.html

  18. [26]

    Scaling language models: Methods, analysis & insights from training gopher,

    J. W. Rae, S. Borgeaud, T. Cai, K. Millican, J. Hoffmann, F. Song, J. Aslanides, S. Henderson, R. Ring, S. Young et al., “Scaling language models: Methods, analysis & insights from training gopher,” arXiv preprint arXiv:2112.11446, 2021

  19. [27]

    How to fine-tune bert for text classification?

    C. Sun, X. Qiu, Y . Xu, and X. Huang, “How to fine-tune bert for text classification?” arXiv preprint arXiv:1905.05583 , 2019

  20. [28]

    Fine- tuning pretrained language models: Weight initializations, data orders, and early stopping,

    J. Dodge, S. Gururangan, D. Card, R. Schwartz, and N. A. Smith, “Fine- tuning pretrained language models: Weight initializations, data orders, and early stopping,” arXiv preprint arXiv:2002.06305 , 2020

  21. [29]

    Revealing the dark secrets of bert,

    O. Kovaleva, A. Romanov, A. Rogers, and A. Rumshisky, “Revealing the dark secrets of bert,” arXiv preprint arXiv:1908.08593 , 2019

  22. [30]

    Amazon Reviews Dataset,

    K. Jain, “Amazon Reviews Dataset,” https://www.kaggle.com/datasets/ kritanjalijain/amazon-reviews/data, 2021

  23. [31]

    Learning word vectors for sentiment analysis,

    A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y . Ng, and C. Potts, “Learning word vectors for sentiment analysis,” in Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , Portland, Oregon, USA, June 2011, pp...

  24. [32]

    Tripadvisor hotel review dataset,

    B. Bansal, “Tripadvisor hotel review dataset,” April 2018, available online: https://zenodo.org/record/1219899

  25. [33]

    Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing,

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig, “Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing,” arXiv preprint arXiv:2107.13586, 2021

  26. [34]

    Making pre-trained language models better few-shot learners,

    T. Gao, A. Fisch, and D. Chen, “Making pre-trained language models better few-shot learners,” in Proceedings of the Association for Compu- tational Linguistics (ACL) , 2021

  27. [35]

    Exploiting cloze questions for few shot text classification and natural language inference,

    T. Schick and H. Sch ¨utze, “Exploiting cloze questions for few shot text classification and natural language inference,” in Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, 2021

  28. [36]

    Google colaboratory,

    Google, “Google colaboratory,” https://colab.research.google.com/, ac- cessed: June 18, 2024

  29. [2023]

    Available: https://arxiv.org/abs/2305.05176

    [Online]. Available: https://arxiv.org/abs/2305.05176

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.