Pith. sign in

REVIEW 1 major objections 5 minor 58 references

MPLinker: Multi-template Prompt-tuning with Adversarial Training for Issue-commit Link Recovery

T0 review · 1 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read MPLinker claims that issue-commit link recovery should be a cloze task, and reports 96.10% average F1 by combining multi-template prompt-tuning with adversarial training.

desk verdict Prompt-tuning for issue-commit links is a reasonable new application, but don't trust the 13-point F1 gap until the baseline protocol is actually shared. read the letter →

arxiv 2501.19026 v1 pith:KYB26RMU submitted 2025-01-31 cs.SE

Bangchao Wang , Yang Deng , Ruiqi Luo , Peng Liang , Tingting Bi This is my paper · ORCID
classification cs.SE
keywords issue-commitlinkrecoveryprompt-tuningmulti-templatepromptingadversarialtrainingpre-trainedlanguagemodelssoftwaretraceabilityclozetaskRoBERTa
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

MPLinker claims that issue-commit link recovery is better framed as a cloze task than as a classification problem bolted onto a pre-trained language model. The method feeds an issue, a commit, and a prompt template containing a [MASK] token into RoBERTa, lets the masked-language-model head predict link-describing words, and averages predictions over several templates. Adversarial training, implemented as a projected-gradient-descent perturbation of the input embedding, is added to reduce overfitting. On six open-source projects the paper reports an average F1 of 96.10% (against 82.77% for the BTLink baseline), with precision 96.49%, recall 95.92%, MCC 94.04%, AUC 96.05%, and accuracy 98.15%. The paper concludes that prompt-tuning is a stronger paradigm for software traceability than the fine-tuning approaches that currently define the state of the art, with direct consequences for bug localization, defect prediction, and security assurance.

What carries the argument

The load-bearing machinery is the prompt template plus verbalizer, used with the MLM head of a pre-trained language model (RoBERTa). The template inserts a [MASK] token after a natural-language phrase; the verbalizer maps the MLM's vocabulary probability distribution to label words ('correct', 'incorrect', 'true', 'false') that stand for linked and not linked. Multi-template prompting averages the label probabilities predicted by each single-template prompt, so no single template choice dominates the outcome. Adversarial training perturbs the input embedding with a projected gradient descent step before the MLM prediction, forcing the model to learn from noisy inputs. This combination is what realigns the downstream task with RoBERTa's masked-language-model pre-training objective, which is the mechanism the paper credits for the performance gain.

What would settle it

Run MPLinker and BTLink on the six datasets under one protocol with the same splits, negative sampling, and number of runs, and compare F1; if BTLink closes most of the 13-point gap or MPLinker drops below roughly 90% F1, the reported advantage would not come from prompt-tuning itself.

Watch

Extended reading notes

Core claim

The central claim is that the mismatch between PLM pre-training and downstream ILR fine-tuning is the main reason existing methods under-use semantic information, and that aligning the downstream task with the pre-training objective fixes it. MPLinker converts link recovery into masked-token prediction: each issue-commit pair is combined with templates such as 'The link is [MASK]', RoBERTa's MLM head assigns vocabulary probabilities, and a verbalizer maps words like 'correct' and 'incorrect' to linked and not-linked. Multiple templates are scored separately and their label probabilities are averaged, which the paper says smooths away the bias of any single template. Adversarial training adds a PGD perturbation to the input embedding before prediction, and the ablation study reports that this improves F1, recall, AUC, and accuracy while improving stability across projects. Compared with FRLink, DeepLink, hybrid-linker, and BTLink, MPLinker achieves the highest average on every metric across the six projects, with large effect sizes, and the paper notes the trade-off on log4net where precision trails BTLink while recall and F1 improve.

Load-bearing premise

The load-bearing premise is that MPLinker and every baseline were evaluated on identical train/validation/test splits and identical negative issue-commit pairs; the paper says the same datasets were used but does not report the baseline re-runs or negative-sampling scheme in enough detail to verify this.

Editorial extensions

If this is right

  • If the reported results are right, ILR systems no longer need task-specific classification heads: a masked-language-model head and a small set of label words are sufficient.
  • Multi-template averaging removes the guesswork of choosing a single prompt, since combining diverse templates yields more stable predictions.
  • Adversarial training at the embedding level acts as a regularizer for ILR, with the paper's ablation showing gains of up to 13.75 F1 points on some projects.
  • The MLM consistency of the PLM with the cloze task matters more than model family, since RoBERTa beats BERT and GPT-2 despite BERT's extra next-sentence prediction objective and GPT-2's autoregressive objective.

Reading between the lines

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

  • The paper does not test this, but the same cloze mechanism could be applied to other traceability link types—for example pull-request-to-issue or requirement-to-code—by swapping the natural-language template and label words.
  • The paper fixes adversarial training to one PGD step (T=1) and a small epsilon; a natural extension would vary the number of steps or perturbation bound to see whether the gains grow or reverse, especially on small projects like log4net.
  • Because the comparison's validity depends on identical splits and negative issue-commit candidates, an independent re-run of MPLinker and BTLink under one shared protocol would clarify how much of the 13-point F1 gap is method and how much is benchmark setup.
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

1 major / 5 minor

Summary. The paper proposes MPLinker, a prompt-tuning approach for issue-commit link recovery that reformulates the binary link prediction task as a masked-language-model cloze task, combines three human-designed templates by averaging their label-word probabilities, and adds adversarial training based on a single-step PGD perturbation. The method is evaluated on six open-source projects (log4net, OODT, Giraph, Keras, Nutch, Isis) with six metrics, three PLMs (RoBERTa, BERT, GPT-2), three prompt architectures, and four baselines (FRLink, DeepLink, hybrid-linker, BTLink). The headline claim, repeated in the abstract, Section 5.4, and Section 7, is that MPLinker achieves average F1 96.10%, Precision 96.49%, Recall 95.92%, MCC 94.04%, AUC 96.05%, and ACC 98.15%, significantly outperforming the state of the art, in particular BTLink, by roughly 13 F1 points.

Significance. If the headline comparison were established under a controlled protocol, the paper would make a useful contribution: it is the first work, to the authors' stated knowledge, to transform ILR into a cloze task, and the multi-template averaging and adversarial-training components are simple and reproducible ideas that could transfer to other software traceability tasks. The paper also ships a replication package and makes an effort to use multiple metrics and non-parametric significance tests, which is commendable. At present, however, the claimed superiority over SOTA rests on an unverified assumption of protocol identity between MPLinker and the baselines, and several mechanical errors in the reported results (an incorrect ACC formula, a duplicated row in Table 3, single-run estimates without variance) weaken the evidence. The central idea remains plausible and testable, but the evidence in the current version is not yet sufficient to support the 'significantly outperforming' claim.

major comments (1)
  1. [Appendix / References] The paper does not report the training time, inference time, or computational cost of MPLinker, which is relevant for a software engineering audience that may want to adopt the method. Adding a brief runtime comparison with the baselines would strengthen the practical value.
minor comments (5)
  1. [§5.4] The phrase 'maintaining their preprocessing methods and data imbalance handling' in Section 5.4 should be made precise: does it mean each baseline used its own preprocessing, or that all methods were unified under MPLinker's preprocessing? This wording is ambiguous and directly affects the interpretation of Table 7.
  2. [§3.2] Equation (16) says '−ϵ is set to 1 by default', which is a sign typo; the intended meaning is that the perturbation bound ϵ is 1. Please correct the notation so readers can reproduce the adversarial setting.
  3. [§2.3] The exposition of CLSPrompt in Section 3.1.3 and Equation (13) is unclear about the role of the [CLS] token: if the input already begins with [CLS], the notation EXprompt = EI ⊕ EC ⊕ ECLS suggests an extra vector is concatenated. Please clarify what is actually concatenated and how the [CLS] output is used for classification.
  4. [Figure 5] The box-plot figure in Section 5.2 has mean and standard-deviation text that is difficult to read in the arXiv version. Please provide a vector figure with legible labels so that the stability claims (e.g., 'Std=4.86') can be verified.
  5. [References] Several references in the related work are informal (e.g., URLs without full bibliographic details) and the text sometimes cites a reference number incorrectly (e.g., 'hybrid-linker [6]' in §5.4 vs. [7] elsewhere). A careful reference cleanup is needed before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MPLinker's reported gains are measured test-set outputs; the multi-template averaging and verbalizer mapping are not fitted to test labels, and no load-bearing claim reduces to its inputs.

full rationale

The paper does not contain a derivation chain in which a prediction is equivalent to its inputs by construction. The verbalizer (Eq. 6) and label-probability mapping (Eq. 7) convert MLM vocabulary probabilities into label-word probabilities, and multi-template averaging (Eq. 12) is an explicit average of per-template probabilities; training uses cross-entropy loss (Eq. 8) on the training split. Nothing in these equations defines the test labels or the reported F1/MCC/AUC values as constructed outputs of the method. The Table 7 numbers are empirical measurements, not fitted parameters renamed as predictions. The choice of multi-template prompt, RoBERTa, and adversarial training was made after observing results on the same six projects (Sections 5.1-5.3), which is a test-set reuse / model-selection threat that can inflate absolute performance, but it is not circular because the final evaluation is still a measurement rather than a reduction to the selection criterion. Similarly, the baseline comparison in Section 5.4 asserts only that the same datasets were used, 'maintaining their preprocessing methods and data imbalance handling'; if the baselines were not re-run under identical splits and negative-candidate sets, that is a protocol-validity concern, not a circularity concern. The only plausible self-citation, reference [55], is used as a stability heuristic (standard deviation below one-tenth of the mean) and is not load-bearing for the central ILR claim. No uniqueness theorem, ansatz-by-citation, or renamed known result appears. The reported result is therefore self-contained as an empirical evaluation, with threats to validity that belong to reproducibility and experimental fairness rather than circularity.

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

MPLinker introduces no new theoretical entities; it combines existing PLMs, prompt templates, and PGD perturbations. The central empirical claims rest on hand-set hyperparameters (epsilon, alpha, T), hand-chosen templates and verbalizer words, and unverified assumptions about benchmark comparability.

free parameters (5)
  • PGD perturbation bound epsilon = 1
    Set by hand in Eq. (16); bounds the L-infinity norm of the embedding perturbation and directly controls the strength of adversarial training that drives the large ablation gains.
  • PGD step size alpha = 1
    Set by hand in Eq. (19); with T=1 it determines the single gradient step applied to the random initial perturbation.
  • PGD iterations T = 1
    Set in Section 3.2 ('we generate the perturbation for each input only once, setting T to 1'), making the adversarial component a one-step noise addition.
  • Verbalizer label word sets = vpos={true, correct, valid, accurate}; vneg={false, wrong, incorrect, invalid}
    Hand-chosen mapping (Eq. 6) between vocabulary words and the binary link label; different label words would change the predicted probability estimates.
  • Prompt templates = Three hand-designed templates (Table 2)
    The template contents and count are chosen without a search; multi-template averaging over these exact templates produces the reported RQ1/RQ4 results.
assumptions (4)
  • domain assumption All methods are evaluated on the same train/valid/test splits and negative candidate sets.
    Section 4.1 and 5.4 state the datasets are from prior work and preprocessing is maintained, but the paper does not demonstrate that baselines were re-run under identical splits and negative sampling; the entire SOTA comparison depends on this.
  • domain assumption Non-linked issue-commit pairs in the benchmark serve as negative samples in the stated counts.
    Section 3 defines false samples as 'links other than cj', yet the reported dataset sizes (e.g., log4net: 239 issues, 115 commits, 866 false links) imply a specific negative-sampling scheme that is not described in the paper.
  • domain assumption MLM-style cloze prompting better exploits RoBERTa's pre-training than fine-tuning-based ILR heads.
    This is the paper's central motivation (Sections 1, 2.2, 5.2), asserted as the reason for the gains, but the experiments do not isolate it from other differences such as model capacity or the adversarial-training component.
  • standard math Wilcoxon signed-rank tests with n=6 datasets and p<0.05 establish significant superiority.
    Section 4.2 adopts the test, but with six projects the smallest achievable two-sided p-value is 0.0313, so all reported significant differences are at the floor value; no correction is made for six metrics and multiple comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MPLinker: Multi-template Prompt-tuning with Adversarial Training for Issue-commit Link Recovery." pith.science (2026). https://pith.science/paper/KYB26RMU

@misc{pith2026250119026,
  author       = {Pith},
  title        = {Pith review of: MPLinker: Multi-template Prompt-tuning with Adversarial Training for Issue-commit Link Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KYB26RMU}},
  note         = {Machine review of arXiv:2501.19026}
}
read the original abstract

In recent years, the pre-training, prompting and prediction paradigm, known as prompt-tuning, has achieved significant success in Natural Language Processing (NLP). Issue-commit Link Recovery (ILR) in Software Traceability (ST) plays an important role in improving the reliability, quality, and security of software systems. The current ILR methods convert the ILR into a classification task using pre-trained language models (PLMs) and dedicated neural networks. these methods do not fully utilize the semantic information embedded in PLMs, resulting in not achieving acceptable performance. To address this limitation, we introduce a novel paradigm: Multi-template Prompt-tuning with adversarial training for issue-commit Link recovery (MPLinker). MPLinker redefines the ILR task as a cloze task via template-based prompt-tuning and incorporates adversarial training to enhance model generalization and reduce overfitting. We evaluated MPLinker on six open-source projects using a comprehensive set of performance metrics. The experiment results demonstrate that MPLinker achieves an average F1-score of 96.10%, Precision of 96.49%, Recall of 95.92%, MCC of 94.04%, AUC of 96.05%, and ACC of 98.15%, significantly outperforming existing state-of-the-art methods. Overall, MPLinker improves the performance and generalization of ILR models, and introduces innovative concepts and methods for ILR. The replication package for MPLinker is available at https://github.com/WTU-intelligent-software-development/MPLinker

Figures

Figures reproduced from arXiv: 2501.19026 by the authors.

Figure 1
Figure 1. The motivation for prompt-tuning PLMs for ILR [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. The overview of MPLinker. Multi-template Prompt on the left and adversarial [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. The comparison of Single-template Prompt, Multi-template Prompt, and [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance comparison of three PLMs across six projects. [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: The mean, standard deviation for the ILR task across three PLMs: RoBERTa, [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 48 canonical work pages

  1. [1]

    J. Lin, Y. Liu, Q. Zeng, M. Jiang, J. Cleland-Huang, Traceability trans- formed: Generating more accurate links with pre-trained bert models, in: 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), IEEE, 2021, pp. 324–335

    work page 2021

  2. [2]

    Rahman, D

    F. Rahman, D. Posnett, I. Herraiz, P. Devanbu, Sample size vs. bias in defect prediction, in: Proceedings of the 2013 9th joint meeting on foundations of software engineering, 2013, pp. 147–157

    work page 2013

  3. [3]

    S. Kim, T. Zimmermann, E. J. Whitehead Jr, A. Zeller, Predicting faults from cached history, in: 29th International Conference on Software Engineering (ICSE’07), IEEE, 2007, pp. 489–498

    work page 2007

  4. [4]

    A. T. Nguyen, T. T. Nguyen, H. A. Nguyen, T. N. Nguyen, Multi- layered approach for recovering links between bug reports and fixes, in: Proceedings of the ACM SIGSOFT 20th International Symposium on the Foundations of Software Engineering, 2012, pp. 1–11

    work page 2012

  5. [5]

    Y. Sun, Q. Wang, Y. Yang, Frlink: Improving the recovery of missing issue-commit links by revisiting file relevance, Information and Software Technology 84 (2017) 33–47

    work page 2017

  6. [6]

    Y. Sun, C. Chen, Q. Wang, B. Boehm, Improving missing issue- commit link recovery using positive and unlabeled data, in: 2017 32nd IEEE/ACM International Conference on Automated Software Engineer- ing (ASE), IEEE, 2017, pp. 147–152

    work page 2017

  7. [7]

    P. R. Mazrae, M. Izadi, A. Heydarnoori, Automated recovery of issue- commit links leveraging both textual and non-textual data, in: 2021 IEEE International Conference on Software Maintenance and Evolution (ICSME), IEEE, 2021, pp. 263–273

    work page 2021

  8. [8]

    H. Ruan, B. Chen, X. Peng, W. Zhao, Deeplink: Recovering issue-commit links based on deep learning, Journal of Systems and Software 158 (2019) 110406

    work page 2019

Show all 58 references
  1. [9]

    R. Xie, L. Chen, W. Ye, Z. Li, T. Hu, D. Du, S. Zhang, Deeplink: A code knowledge graph based deep learning approach for issue-commit link recovery, in: 2019 IEEE 26th International Conference on Software 32 Analysis, Evolution and Reengineering (SANER), IEEE, 2019, pp. 434– 444

    2019

  2. [10]

    S. Bai, L. Liu, H. Liu, M. Zhang, C. Meng, P. Zhang, Find potential partners: A github user recommendation method based on event data, Inf. Softw. Technol. 150 (C) (Oct. 2022)

    2022

  3. [11]

    S. Bai, H. Liu, E. Dai, L. Liu, Improving issue-pr link prediction via knowledge-aware heterogeneous graph learning, IEEE Transactions on Software Engineering 50 (7) (2024) 1901–1920

    2024

  4. [12]

    Batra, L

    H. Batra, L. Nelson, Esd: E-mail spam detection using cybersecurity- driven header analysis and machine learning based content analysis., International Journal of Performability Engineering 20 (4) (2024)

    2024

  5. [13]

    Nandwalkar, S

    B. Nandwalkar, S. Pardeshi, M. Shahade, A. Awate, Descriptive hand- written paper grading system using nlp and fuzzy logic, International Journal of Performability Engineering 19 (4) (2023) 273

    2023

  6. [14]

    J. Lan, L. Gong, J. Zhang, H. Zhang, Btlink: automatic link recovery between issues and commits based on pre-trained bert model, Empirical Software Engineering 28 (4) (2023) 103

    2023

  7. [15]

    Zhang, Y

    C. Zhang, Y. Wang, Z. Wei, Y. Xu, J. Wang, H. Li, R. Ji, Ealink: An efficient and accurate pre-trained framework for issue-commit link recovery, ArXiv abs/2308.10759 (2023). URL https://api.semanticscholar.org/CorpusID:261049544

    2023 arXiv

  8. [16]

    Zhang, B

    Z. Zhang, B. wei Wang, Prompt learning for news recommendation, Pro- ceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (2023). URL https://api.semanticscholar.org/CorpusID:258059770

    2023

  9. [17]

    Y. Gu, X. Han, Z. Liu, M. Huang, Ppt: Pre-trained prompt tuning for few-shot learning, arXiv preprint arXiv:2109.04332 (2021)

    2021 arXiv

  10. [18]

    Z. Yang, S. Wang, B. P. S. Rawat, A. Mitra, H. Yu, Knowledge in- jected prompt based fine-tuning for multi-label few-shot icd coding, in: Proceedings of the Conference on Empirical Methods in Natural Lan- guage Processing. Conference on Empirical Methods in Natural Language Pr...

    2022

  11. [19]

    X. Chen, N. Zhang, X. Xie, S. Deng, Y. Yao, C. Tan, F. Huang, L. Si, H. Chen, Knowprompt: Knowledge-aware prompt-tuning with synergistic optimization for relation extraction, in: Proceedings of the ACM Web conference 2022, 2022, pp. 2778–2788

    2022

  12. [20]

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, G. Neubig, Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, ACM Computing Surveys 55 (9) (2023) 1–35

    2023

  13. [21]

    J. H. Hayes, A. Dekhtyar, S. K. Sundaram, Advancing candidate link generation for requirements tracing: The study of methods, IEEE Trans- actions on Software Engineering 32 (1) (2006) 4–19

    2006

  14. [22]

    De Lucia, F

    A. De Lucia, F. Fasano, R. Oliveto, G. Tortora, Enhancing an artefact management system with traceability recovery features, in: 20th IEEE International Conference on Software Maintenance, 2004. Proceedings., IEEE, 2004, pp. 306–315

    2004

  15. [23]

    H. Gao, H. Kuang, K. Sun, X. Ma, A. Egyed, P. M¨ ader, G. Rong, D. Shao, H. Zhang, Using consensual biterms from text structures of requirements and code to improve ir-based traceability recovery, in: Proceedings of the 37th IEEE/ACM International Conference on Automated Softw...

    2022

  16. [24]

    Mahmoud, N

    A. Mahmoud, N. Niu, Supporting requirements traceability through refactoring, in: 2013 21st IEEE International Requirements Engineering Conference (RE), IEEE, 2013, pp. 32–41

    2013

  17. [25]

    H. Gao, H. Kuang, X. Ma, H. Hu, J. L¨ u, P. M¨ ader, A. Egyed, Propagating frugal user feedback through closeness of code dependencies to improve ir-based traceability recovery, Empirical Software Engineering 27 (2) (2022) 41

    2022

  18. [26]

    D. V. Rodriguez, D. L. Carver, An ir-based artificial bee colony approach for traceability link recovery, in: 2020 IEEE 32nd International Con- ference on Tools with Artificial Intelligence (ICTAI), IEEE, 2020, pp. 1145–1153

    2020

  19. [27]

    D. V. Rodriguez, D. L. Carver, Multi-objective information retrieval- based nsga-ii optimization for requirements traceability recovery, in: 2020 34 IEEE International Conference on Electro Information Technology (EIT), IEEE, 2020, pp. 271–280

    2020

  20. [28]

    Du, G.-h

    T.-b. Du, G.-h. Shen, Z.-q. Huang, Y.-s. Yu, D.-x. Wu, Automatic traceability link recovery via active learning, Frontiers of Information Technology & Electronic Engineering 21 (8) (2020) 1217–1225

    2020

  21. [29]

    L. Dong, H. Zhang, W. Liu, Z. Weng, H. Kuang, Semi-supervised pre- processing for learning-based traceability framework on real-world soft- ware projects, in: Proceedings of the 30th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Softwar...

    2022

  22. [30]

    Mills, J

    C. Mills, J. Escobar-Avila, A. Bhattacharya, G. Kondyukov, S. Chakraborty, S. Haiduc, Tracing with less data: active learning for classification-based traceability link recovery, in: 2019 IEEE Inter- national Conference on Software Maintenance and Evolution (ICSME), IEEE, 2019...

    2019

  23. [31]

    Moran, D

    K. Moran, D. N. Palacio, C. Bernal-C´ ardenas, D. McCrystal, D. Poshy- vanyk, C. Shenefiel, J. Johnson, Improving the effectiveness of traceability link recovery using hierarchical bayesian networks, in: Proceedings of the ACM/IEEE 42nd International Conference on Software Eng...

    2020

  24. [32]

    Mikolov, I

    T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, J. Dean, Distributed representations of words and phrases and their compositionality, Ad- vances in neural information processing systems 26 (2013)

    2013

  25. [33]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural compu- tation 9 (8) (1997) 1735–1780

    1997

  26. [34]

    J. Guo, J. Cheng, J. Cleland-Huang, Semantically enhanced software traceability using deep learning techniques, in: 2017 IEEE/ACM 39th International Conference on Software Engineering (ICSE), IEEE, 2017, pp. 3–14

    2017

  27. [35]

    Z. Feng, D. Guo, D. Tang, N. Duan, X. Feng, M. Gong, L. Shou, B. Qin, T. Liu, D. Jiang, et al., Codebert: A pre-trained model for programming and natural languages, arXiv preprint arXiv:2002.08155 (2020). 35

    2020 arXiv

  28. [37]

    X. L. Li, P. Liang, Prefix-tuning: Optimizing continuous prompts for generation, In Proceedings of ACL/IJCNLP 2021 (2021)

    2021

  29. [38]

    T. Gao, A. Fisch, D. Chen, Making pre-trained language models better few-shot learners, In Proceedings of ACL (2021)

    2021

  30. [39]

    T. Shin, Y. Razeghi, R. L. Logan IV, E. Wallace, S. Singh, Autoprompt: Eliciting knowledge from language models with automatically generated prompts, In Proceedings of EMNLP 2020 (2020)

    2020

  31. [40]

    J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, G. Neubig, Towards a unified view of parameter-efficient transfer learning, arXiv preprint arXiv:2110.04366 (2021)

    2021 arXiv

  32. [41]

    Hambardzumyan, H

    K. Hambardzumyan, H. Khachatrian, J. May, Warp: Word-level adver- sarial reprogramming, In Proceedings of ACL/IJCNLP 2021 (2021)

    2021

  33. [42]

    C. Zhu, Y. Cheng, Z. Gan, S. Sun, T. Goldstein, J. Liu, Freelb: Enhanced adversarial training for natural language understanding, arXiv preprint arXiv:1909.11764 (2019)

    2019 arXiv

  34. [43]

    Bayer, M.-A

    M. Bayer, M.-A. Kaufhold, C. Reuter, A survey on data augmentation for text classification, ACM Computing Surveys 55 (7) (2022) 1–39

    2022

  35. [44]

    Geisler, T

    S. Geisler, T. Wollschl¨ ager, M. Abdalla, J. Gasteiger, S. G¨ unnemann, Attacking large language models with projected gradient descent, arXiv preprint arXiv:2402.09154 (2024)

    2024 arXiv

  36. [45]

    Moothedath, N

    S. Moothedath, N. Vaswani, Comparing decentralized gradient descent approaches and guarantees, in: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023, pp. 1–5

    2023

  37. [46]

    Zhang, H

    M. Zhang, H. Liu, C. Chen, Y. Liu, S. Bai, Consistent or not? an investigation of using pull request template in github, Information and Software Technology 144 (2022) 106797. 36

    2022

  38. [47]

    S. Bai, L. Liu, C. Meng, H. Liu, Automating discussion structure re- organization for github issues, Expert Syst. Appl. 225 (2023) 120024

    2023

  39. [48]

    R. F. Woolson, Wilcoxon signed-rank test, Encyclopedia of Biostatistics 8 (2005)

    2005

  40. [49]

    Macbeth, E

    G. Macbeth, E. Razumiejczyk, R. D. Ledesma, Cliff’s delta calculator: A non-parametric effect size program for two groups of observations, Universitas Psychologica 10 (2) (2011) 545–555

    2011

  41. [50]

    Kitani, H

    M. Kitani, H. Murakami, One-sample location test based on the sign and wilcoxon signed-rank tests, Journal of Statistical Computation and Simulation 92 (3) (2022) 610–622

    2022

  42. [51]

    Meissel, E

    K. Meissel, E. S. Yao, Using cliff’s delta as a non-parametric effect size measure: An accessible web app and r tutorial, Practical Assessment, Research, and Evaluation 29 (1) (2024)

    2024

  43. [52]

    Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized BERT pretraining approach, CoRR abs/1907.11692 (2019). arXiv:1907.11692. URL http://arxiv.org/abs/1907.11692

    2019 arXiv

  44. [53]

    Devlin, M

    J. Devlin, M. Chang, K. Lee, K. Toutanova, BERT: pre-training of deep bidirectional transformers for language understanding, CoRR abs/1810.04805 (2018). arXiv:1810.04805. URL http://arxiv.org/abs/1810.04805

    2018 arXiv

  45. [54]

    Radford, J

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, Language models are unsupervised multitask learners (2019)

    2019

  46. [55]

    B. Wang, R. Peng, Z. Wang, X. Wang, Y. Li, An automated hybrid approach for generating requirements trace links, Int. J. Softw. Eng. Knowl. Eng. 30 (2020) 1005–1048. URL https://api.semanticscholar.org/CorpusID:221743812

    2020

  47. [56]

    T.-D. B. Le, M. Linares-V´ asquez, D. Lo, D. Poshyvanyk, Rclinker: Automated linking of issue reports and commits leveraging rich contextual information, in: 2015 IEEE 23rd international conference on program comprehension, IEEE, 2015, pp. 36–47. 37

    2015

  48. [57]

    Y. Qu, P. Liu, W. Song, L. Liu, M. Cheng, A text generation and prediction system: pre-training on new corpora using bert and gpt-2, in: 2020 IEEE 10th international conference on electronics information and emergency communication (ICEIEC), IEEE, 2020, pp. 323–326

    2020

  49. [58]

    Demırcı, C

    D. Demırcı, C. Acarturk, et al., Static malware detection using stacked bilstm and gpt-2, IEEE Access 10 (2022) 58488–58502

    2022

  50. [59]

    Lajk´ o, V

    M. Lajk´ o, V. Csuvik, L. Vid´ acs, Towards javascript program repair with generative pre-trained transformer (gpt-2), in: Proceedings of the Third International Workshop on Automated Program Repair, 2022, pp. 61–68. 38

    2022

Pith tools

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