Pith. sign in

REVIEW 4 major objections 6 minor 42 references

Modeling Data Diversity for Joint Instance and Verbalizer Selection in Cold-Start Scenarios

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that in cold-start prompt classification the choices of which instances to annotate and which vocabulary tokens to use as verbalizers should be made jointly, and it reports accuracy gains over separate selection on eight…

desk verdict A clever idea undone by a sign error: Eq. (9) defines separation as the maximum cosine similarity to another cluster, so maximizing it rewards overlapping clusters, directly contradicting the paper's diversity claim—but the joint-selection idea and broad experiments still justify a serious referee. read the letter →

arxiv 2507.00330 v1 pith:TQI3PBMQ submitted 2025-07-01 cs.CL cs.IR

classification cs.CLcs.IR
keywords cold-startsettingprompt-basedlearningverbalizerselectionfew-shotinstancedatadiversitymodelingactivemaskedlanguagemodels
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, in cold-start prompt-based classification (no labeled data at the start), the choice of which unlabeled instances to annotate and the choice of which vocabulary words serve as class labels (verbalizers) should be made jointly, because an instance's label probability depends on how close a verbalizer token is to its masked-token embedding. It proposes COLDSELECT, which projects vocabulary-token embeddings and $h_{[\mathrm{MASK}]}$ embeddings into one space, clusters them with KMeans after PCA and silhouette-score refinement, and then spends a labeling budget by repeatedly selecting clusters that maximize a score built from cohesion, separation, and impurity. The paper says this is the first joint method for these two selections, and on eight benchmark datasets it reports higher accuracy than instance-selection baselines like PATRON, verbalizer baselines like LM-BFF and ProtoVerb, and random selection. If the result holds, prompt-based few-shot classifiers would become less sensitive to arbitrary choices of examples and label words, and annotation budgets would be spent more efficiently.

What carries the argument

The central object is a shared embedding space plus a cluster-score selection loop. The space makes pre-softmax vocabulary token embeddings and the dataset instances' $h_{[\mathrm{MASK}]}$ embeddings directly comparable; PCA reduces their dimension, KMeans partitions them, and a negative silhouette loss refines the partition into cohesive, separated clusters. Cluster refinement discards token-only clusters and reassigns instance-only clusters to the nearest mixed cluster. The selection loop then scores each cluster by cohesion (mean similarity to its verbalizer tokens), separation (similarity to the nearest verbalizer token outside the cluster), and impurity (label diversity), and repeatedly chooses the cluster with the highest score until the budget is spent. Instance and verbalizer choices follow from Eqs. (14)--(16), so the two selections share one budget and one objective.

What would settle it

Rerun COLDSELECT on, say, IMDB with the same budget but replace the verbalizer token chosen by Eq. (15) with a random vocabulary token from the same cluster; if accuracy stays flat, the claim that joint verbalizer selection drives the gains is falsified. A second check is to inspect the chosen tokens and measure whether they are semantically aligned with the class labels they are meant to name.

Watch

Extended reading notes

Core claim

The central claim is that instance selection and verbalizer selection are one problem, not two. The paper grounds this in Eq. (1), where the probability of label $y$ for instance $I$ is a softmax over dot products between the verbalizer token embedding $w_{M(y)}$ and the masked-token embedding $h_{[\mathrm{MASK}]}$; a verbalizer token that lies close to an instance's embedding makes that label likely. COLDSELECT operationalizes that coupling by clustering both kinds of embeddings in a shared space, then spending the labeling budget on clusters that maximize cohesion plus separation plus impurity. For each chosen cluster, it labels the most central instance and adopts the nearest vocabulary token as the verbalizer for that instance's label. The reported experiments, on SST-2, MR, CR, Subj, CoLA, AG News, Yelp-full, and IMDB, show accuracy gains over instance-selection baselines such as PATRON and verbalizer baselines such as LM-BFF and ProtoVerb.

Load-bearing premise

The load-bearing premise is that the nearest vocabulary token in the shared embedding space carries the class label of the just-annotated instance; if that token is a stopword or semantically unrelated, the verbalizer misrepresents the class and the joint-selection gains collapse.

Editorial extensions

If this is right

  • Cold-start prompt classification would no longer need manually designed verbalizers: the label words are selected in the same pass as the examples to annotate.
  • The same labeling budget buys more signal, since COLDSELECT reports needing a smaller budget than random and random-g selection in the RoBERTa-large experiments.
  • Datasets with many classes and imbalanced distributions should show the largest gains; the paper's biggest margins are on AG News and Yelp-full.
  • Prompt-based few-shot classifiers would become less sensitive to arbitrary seed examples, because both the seed set and the label vocabulary are chosen by the same uncertainty-reduction criterion.

Reading between the lines

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

  • A natural extension not tested in the paper is a full active-learning loop that re-clusters after each batch; the dynamic cohesion and separation updates suggest the method is built for that.
  • The quality of Eq. (15)'s verbalizer tokens can be audited directly: list the chosen tokens and check semantic alignment with class names, a check the paper does not report.
  • Replacing PCA with a nonlinear reduction is a testable variant the paper itself flags as future work, and would show how much the result depends on the linear-separability assumption.
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 / 6 minor

Summary. The paper proposes COLD SELECT, a cold-start method that jointly selects few-shot instances to annotate and verbalizer tokens for prompt-based classification with masked language models. The method maps vocabulary embeddings and h[MASK] instance embeddings into a shared space, applies PCA, KMeans clustering, and negative silhouette loss, then iteratively selects clusters and instances under a labeling budget using cohesion, separation, and impurity terms. Experiments on eight benchmarks compare COLD SELECT with instance-selection baselines and verbalizer-selection baselines, reporting consistent accuracy improvements. The paper claims to be the first joint instance and verbalizer selection method in cold-start settings and to reduce labeling uncertainty by modeling data diversity.

Significance. If the technical issues are resolved, this would be a useful contribution to few-shot prompt-based learning: it releases code, evaluates on eight benchmarks, and addresses an underexplored coupling between instance annotation and verbalizer choice. The proposed problem formulation is timely, and the comparison set is broad. However, the current manuscript contains a central inconsistency in the separation term of the selection objective and an underspecified verbalizer aggregation rule, so the empirical gains cannot yet be attributed to the stated mechanism.

major comments (4)
  1. [Section 4.4, Eqs. (9), (11), (13)] The separation term is defined as the maximum cosine similarity to another cluster centroid, which measures proximity, not distance, and Eq. (11) maximizes this quantity. This means the selection objective prefers clusters that are most similar to another cluster, the opposite of the stated goal of avoiding redundancy and capturing inter-cluster distinctiveness. The same sign problem appears in the dynamic update, Eq. (13), which maximizes the maximum cosine similarity to out-of-cluster verbalizer tokens. Unless the implementation uses a decreasing function such as 1 minus the maximum similarity, the optimization does not model separation at all, and the 'Cohesion + Separation' ablation row in Table 5 does not test the stated separation concept.
  2. [Section 4.4, Eqs. (12)-(13) and Eq. (15)] The dynamic cohesion and separation updates are not well-defined because the symbol V is overloaded. In Eq. (1), V denotes the PLM vocabulary, while in Eq. (15) the selected verbalizer tokens are added to V. If V is the full vocabulary, then Eq. (12) maximizes over all vocabulary tokens in a cluster, not over selected class verbalizers, so it does not implement the stated replacement of static centroids by verbalizer-token embeddings. If V is the set of selected verbalizer tokens, then it is empty at initialization, making Eqs. (12)-(13) undefined until the first selection. The paper must define a single consistent notation and specify how the selected tokens are used.
  3. [Section 4.4, Eq. (15) and Eq. (1)] The paper never specifies how multiple verbalizer tokens added to V are aggregated into the label probability p(y|I) in Eq. (1), which is written for a single token M(y). The method may select several tokens for the same class, but no sum, average, or other combination rule is given. Without this aggregation rule, the classification procedure used to produce the experimental results is underspecified.
  4. [Section 4.4, Eq. (15)] The verbalizer selection rule assumes that the nearest vocabulary token to a labeled instance's h[MASK] embedding carries that instance's class label. This assumption is not validated: the nearest token could be a stopword or a semantically unrelated token, degrading classification. The paper should report examples of selected verbalizer tokens or include an ablation that replaces automatically selected verbalizers with manual verbalizers to isolate their contribution.
minor comments (6)
  1. [Section 5.3] The experiments use a single random seed (42) and report no variance; please report means and standard deviations over multiple seeds for the main tables.
  2. [Tables 2 and 3] The column header 'Uncertainity' is misspelled and should read 'Uncertainty'.
  3. [Eq. (3)] The PCA objective should specify that the data matrix X is centered; otherwise the projection is not the standard PCA formulation.
  4. [Figure 1] The figure labels 'Negative Instances' and 'Positive Instances' before any labels exist in the cold-start setting; this is confusing and should be clarified as cluster assignments or predicted groupings.
  5. [Section 4.5] The section is titled 'Optimal Selection Process' and uses the word 'optimal' repeatedly, but no formal optimality proof or well-defined global objective is given; 'greedy' would be more accurate unless a proof is added.
  6. [Section 4.3, Eqs. (4)-(9)] KMeans and the silhouette score are formulated with Euclidean distances, while cohesion and separation in Eqs. (8)-(9) use cosine similarity; the inconsistency between distance metrics should be discussed or reconciled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: COLDSELECT's selection procedure is a constructive heuristic evaluated on held-out benchmarks, not a derivation that reduces to its own inputs.

full rationale

COLDSELECT proposes a concrete pipeline (PCA, KMeans, negative silhouette refinement, then cluster selection via Eqs. 8-16) and tests accuracy on held-out test sets. Verbalizer tokens are chosen as the nearest vocabulary token to a human-labeled instance (Eq. 15), and final accuracy is measured on separate test data; consequently, the reported predictions are not fitted values renamed as results. The method's few self-citations (e.g., [2], [3], [11]) are background references to the authors' prior work and are not load-bearing for the claimed contribution. Section 7 honestly acknowledges limitations of PCA and KMeans, which are standard heuristic concerns rather than circular support. One notable non-circular issue is that Eq. (9) defines separation as the maximum cosine similarity to another cluster while Eq. (11) maximizes that quantity, meaning the term named 'separation' actually rewards inter-cluster proximity; this is a correctness/sign-error concern about the heuristic, not an equation-level reduction of a prediction to its inputs. Overall, the derivation chain is self-contained and not circular.

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

The method relies on the assumptions that embedding-space proximity reflects label semantics, that cluster impurity from acquired labels approximates uncertainty for all instances, and that KMeans on PCA embeddings captures class-relevant diversity. No new physical or conceptual entities are introduced.

free parameters (4)
  • number of clusters K = 40
    Set to 40 for all datasets without sensitivity analysis (Section 5.3).
  • cluster optimization iterations = 5
    Set to 5 (Section 5.3).
  • random seed = 42
    Set for KMeans (Section 5.3).
  • PCA dimensions
    Not specified in Section 5.3; the number of PCA components could affect results.
assumptions (3)
  • domain assumption Pre-softmax token embeddings and [MASK] embeddings share a metric space where cosine similarity reflects semantic and class proximity.
    Used throughout Sections 4.2-4.4 to select verbalizers and instances based on cosine similarity.
  • domain assumption Cluster impurity computed from acquired labels approximates classification uncertainty for all instances in the cluster.
    Section 4.4 equates cluster-level impurity minimization with instance-level classification uncertainty minimization.
  • domain assumption KMeans clustering on PCA-reduced embeddings captures class-relevant diversity in the data.
    Section 4.3 relies on this assumption for the cluster creation step that drives all subsequent selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modeling Data Diversity for Joint Instance and Verbalizer Selection in Cold-Start Scenarios." pith.science (2026). https://pith.science/paper/TQI3PBMQ

@misc{pith2026250700330,
  author       = {Pith},
  title        = {Pith review of: Modeling Data Diversity for Joint Instance and Verbalizer Selection in Cold-Start Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQI3PBMQ}},
  note         = {Machine review of arXiv:2507.00330}
}
abstract

Prompt-based methods leverage the knowledge of pre-trained language models (PLMs) trained with a masked language modeling (MLM) objective; however, these methods are sensitive to template, verbalizer, and few-shot instance selection, particularly in cold-start settings with no labeled data. Existing studies overlook the dependency between instances and verbalizers, where instance-label probabilities depend on verbalizer token proximity in the embedding space. To address this, we propose COLDSELECT, a joint verbalizer and instance selection approach that models data diversity. COLDSELECT maps PLM vocabulary and $h_{[MASK]}$ embeddings into a shared space, applying dimensionality reduction and clustering to ensure efficient and diverse selection. By optimizing for minimal uncertainty and maximal diversity, COLDSELECT captures data relationships effectively. Experiments on eight benchmarks demonstrate COLDSELECT's superiority in reducing uncertainty and enhancing generalization, outperforming baselines in verbalizer and few-shot instance selection for cold-start scenarios.

Figures

Figures reproduced from arXiv: 2507.00330 by the authors.

Figure 1
Figure 1. Overview of COLDSELECT lack a cohesive framework to simultaneously address data diversity and labeling uncertainty, leaving significant room for improvement. Verbalizer Selection. Verbalizers play a crucial role in mapping model predictions to class labels. Early approaches [28,29] relied on manually designed verbalizers, which, while effective, were time-consuming and susceptible to human bias. To automate this pro… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

  2. [2]

    In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Chakraborty, M., Kulkarni, A., Li, Q.: Open-domain aspect-opinion co-mining with double- layer span extraction. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. p. 66–75. KDD ’22, Association for Computing Machinery, New York, NY , USA (2022). https://doi.org/10.1145/3534678.3539386, https://doi.org/10.1145/3534678.3539386

  3. [3]

    In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers)

    Chakraborty, M., Kulkarni, A., Li, Q.: Zero-shot approach to overcome perturbation sensitivity of prompts. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). pp. 5698–5711 (2023)

  4. [4]

    In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 2: Short Papers)

    Chang, E., Shen, X., Yeh, H.S., Demberg, V .: On training instance selection for few-shot neural text generation. In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 2: Short Papers). pp. 8–13 (2021) 14 Mohna Adithya et al

  5. [5]

    In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers)

    Cui, G., Hu, S., Ding, N., Huang, L., Liu, Z.: Prototypical verbalizer for prompt-based few-shot tuning. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). pp. 7014–7024 (2022)

  6. [6]

    In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 1: Long Papers)

    Gao, T., Fisch, A., Chen, D.: Making pre-trained language models better few-shot learners. In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 1: Long Papers). pp. 3816–3830 (2021)

  7. [7]

    In: International Conference on Machine Learning

    Hacohen, G., Dekel, A., Weinshall, D.: Active learning on a budget: Opposite strategies suit high and low budgets. In: International Conference on Machine Learning. pp. 8175–8195. PMLR (2022)

  8. [8]

    In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 1: Long Papers)

    Hambardzumyan, K., Khachatrian, H., May, J.: Warp: Word-level adversarial reprogramming. In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 1: Long Papers). pp. 4921–4933 (2021)

Show all 42 references
  1. [9]

    In: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining

    Hu, M., Liu, B.: Mining and summarizing customer reviews. In: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 168– 177 (2004)

  2. [10]

    In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers)

    Hu, S., Ding, N., et al.: Knowledgeable prompt-tuning: Incorporating knowledge into prompt verbalizer for text classification. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). pp. 2225–2240 (2022)

  3. [11]

    In: Evans, R.J., Shpitser, I

    Kulkarni, A., Chakraborty, M., Xie, S., Li, Q.: Optimal budget allocation for crowdsourcing labels for graphs. In: Evans, R.J., Shpitser, I. (eds.) Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence. Proceedings of Machine Learning Research, v...

  4. [12]

    In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

    Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (2021)

  5. [13]

    In: Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval

    Lewis, D.D., Gale, W.A.: A sequential algorithm for training text classifiers. In: Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval. pp. 3–12 (1994)

  6. [14]

    In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 1: Long Papers)

    Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Proceedings of the 59th Annual Meeting of the ACL and the 11th IJCNLP (V olume 1: Long Papers). pp. 4582–4597 (2021)

  7. [15]

    In: CCF International Conference on Natural Language Processing and Chinese Computing

    Liu, C., Wang, H., Xi, N., Zhao, S., Qin, B.: Global prompt cell: A portable control module for effective prompt tuning. In: CCF International Conference on Natural Language Processing and Chinese Computing. pp. 657–668 (2023)

  8. [16]

    Liu, J., Shen, D., et al.: What makes good in-context examples for gpt-3? In: Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures. pp. 100–114 (2022)

  9. [17]

    AI Open (2023)

    Liu, X., Zheng, Y ., Du, Z., et al.: Gpt understands, too. AI Open (2023)

  10. [18]

    arXiv preprint arXiv:1907.11692 (2019)

    Liu, Y ., Ott, M., et al.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)

  11. [19]

    In: Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies

    Maas, A., Daly, R.E., et al.: Learning word vectors for sentiment analysis. In: Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies. pp. 142–150 (2011)

  12. [20]

    In: Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press (1967)

    MacQueen, J.: Some methods for classification and analysis of multivariate observations. In: Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press (1967)

  13. [21]

    In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

    Margatina, K., Vernikos, G., Barrault, L., Aletras, N.: Active learning by acquiring contrastive examples. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 650–663 (2021)

  14. [22]

    In: Proceedings of the AAAI conference on artificial intelligence

    Meng, Y ., Shen, J., Zhang, C., Han, J.: Weakly-supervised hierarchical text classification. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 6826–6833 (2019) Title Suppressed Due to Excessive Length 15

  15. [23]

    In: International Conference on Applications of Natural Language to Information Systems

    Müller, T., Pérez-Torró, G., Basile, A., et al.: Active few-shot learning with fasl. In: International Conference on Applications of Natural Language to Information Systems. pp. 98–110 (2022)

  16. [24]

    In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2002 (2002)

    PANG, B.: Thumbs up? sentiment classification using machine learning techniques. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2002 (2002)

  17. [25]

    In: Proceedings of the 42nd Meeting of the Association for Computational Linguistics (ACL), 2004 (2004)

    PANG, B.: A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In: Proceedings of the 42nd Meeting of the Association for Computational Linguistics (ACL), 2004 (2004)

  18. [26]

    Journal of computational and applied mathematics 20, 53–65 (1987)

    Rousseeuw, P.J.: Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics 20, 53–65 (1987)

  19. [27]

    In: Proceedings of the 28th International Conference on Computational Linguistics

    Schick, T., Schmid, H., Schütze, H.: Automatically identifying words that can serve as labels for few-shot text classification. In: Proceedings of the 28th International Conference on Computational Linguistics. pp. 5569–5578 (2020)

  20. [28]

    In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main V olume

    Schick, T., Schütze, H.: 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: Main V olume. pp. 255–269 (2021)

  21. [29]

    In: Proceedings of the 2021 NAACL-HLT

    Schick, T., Schütze, H.: It’s not just size that matters: Small language models are also few- shot learners. In: Proceedings of the 2021 NAACL-HLT. pp. 2339–2352 (2021)

  22. [30]

    In: Findings of the ACL: ACL 2022

    Schröder, C., Niekler, A., Potthast, M.: Revisiting uncertainty-based query strategies for active learning with transformers. In: Findings of the ACL: ACL 2022. pp. 2194–2203 (2022)

  23. [31]

    In: International Conference on Learning Representations (2018)

    Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. In: International Conference on Learning Representations (2018)

  24. [32]

    In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Shin, T., Razeghi, Y ., et al.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 4222–4235 (2020)

  25. [33]

    In: Proceedings of the 2013 conference on empirical methods in natural language processing

    Socher, R., Perelygin, A., et al.: Recursive deep models for semantic compositionality over a sentiment treebank. In: Proceedings of the 2013 conference on empirical methods in natural language processing. pp. 1631–1642 (2013)

  26. [34]

    arXiv preprint arXiv:2209.01975 (2022)

    Su, H., Kasai, J., et al.: Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975 (2022)

  27. [35]

    In: Proceedings of the 29th International Conference on Computational Linguistics

    Wang, H., Liu, C., et al.: Prompt combines paraphrase: Teaching pre-trained models to understand rare biomedical words. In: Proceedings of the 29th International Conference on Computational Linguistics. pp. 1422–1431 (2022)

  28. [36]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, H., Zhao, S., et al.: Manifold-based verbalizer space re-embedding for tuning- free prompt-based classification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19126–19134 (2024)

  29. [37]

    Transactions of the Association for Computational Linguistics 7, 625–641 (2019)

    Warstadt, A., Singh, A., Bowman, S.R.: Neural network acceptability judgments. Transactions of the Association for Computational Linguistics 7, 625–641 (2019)

  30. [38]

    Chemometrics and intelligent laboratory systems 2(1-3), 37–52 (1987)

    Wold, S., Esbensen, K., Geladi, P.: Principal component analysis. Chemometrics and intelligent laboratory systems 2(1-3), 37–52 (1987)

  31. [39]

    In: Proceedings of the 61st Annual Meeting of the ACL (V olume 1: Long Papers)

    Yu, Y ., Zhang, R., et al.: Cold-start data selection for better few-shot language model fine-tuning: A prompt-based uncertainty propagation approach. In: Proceedings of the 61st Annual Meeting of the ACL (V olume 1: Long Papers). pp. 2499–2521 (2023)

  32. [40]

    In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Yuan, M., Lin, H.T., Boyd-Graber, J.: Cold-start active learning through self-supervised language modeling. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 7935–7948 (2020)

  33. [41]

    In: International Conference on Learning Representations

    Zhang, N., Li, L., et al.: Differentiable prompt makes pre-trained language models better few-shot learners. In: International Conference on Learning Representations

  34. [42]

    Advances in neural information processing systems 28 (2015)

    Zhang, X., Zhao, J., LeCun, Y .: Character-level convolutional networks for text classification. Advances in neural information processing systems 28 (2015)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.