9.5 KiB
9.5 KiB
type:: REVIEWS tags:: year:: 2024 venue:: ESEM full-title:: PromptLink: Multi-template prompt learning with adversarial training for issue-commit link recovery date-start:: 15-07-2024 - 11:49 date-submitted:: 15-07-2024 external-links:: status:: DONE deadline-submission:: 16-07-2024 file:: @ESEM24_paper_220.pdf parent:: todoist:: https://app.todoist.com/app/task/esem-24-submission-assignment-6VVGWMVqvVwcp598
- ### [[Highlights]]
- # Annotazioni
(15/7/2024, 15:59:52)
- “issue-commit link recovery (ILR)” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=G6IIJTEF)) #5fb236
- “due to inconsistencies between the ILR task and PLMs, these methods not fully leverage the semantic information in PLMs.” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=UX57XRQ9)) #5fb236
- “Multi-template prompt learning method with adversarial training for issue-commit link recovery” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=MBF6CUCN)) #ffd400
*What does it mean?*
- “trace links between issues and commits (issue-commit link recovery (ILR)) play a key role in bug localization[1], bug prediction[2], and other software maintenance tasks.” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=EYLB3BNC)) #a28ae5
- “However, these methods require large-scale data and not apply to projects with poor quality or small scales.” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=7UB7GDZB)) #5fb236
- “Although these methods demonstrate good performance, they are not consistent with the training objectives of PLMs, which may result in generalized domain knowledge in PLMs not being well used.” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=6BE62UYL)) #a28ae5
- “prompt learning” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=TSDJ69GJ)) #e56eee
- “prompting templates and label words” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=MT7UYXNK)) #5fb236
- “Multi-template prompt learning method” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=9CRGFGAR)) #5fb236
- “adversarial training” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=KYCDLFK7)) #ffd400
*why? What's the role of the adversarial training?*
- “mitigate the model overfitting” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=YC8XUCUP)) #5fb236
- “Single and Multi-template PromptLink for the ILR?” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=DTFECZKU)) #5fb236
- “adversarial training contribute to the PromptLink” ([“ESEM24_paper_220.pdf”, p. 1](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=1&annotation=AKEXKUI6)) #5fb236
- “PromptLink significantly enhances ILR performance, revealing the potential of prompt learning in ILR tasks.” ([“ESEM24_paper_220.pdf”, p. 2](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=2&annotation=Z5YXRUZ9)) #a28ae5
- “However, these approaches necessitate extensive feature engineering, and the feature selection needs researchers to have a substantial reserve of prior knowledge.” ([“ESEM24_paper_220.pdf”, p. 2](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=2&annotation=JNW2HY8H)) #5fb236
- “minimize the need for manual feature engineering and leverage more automated techniques to enhance the performance of the ILR model.” ([“ESEM24_paper_220.pdf”, p. 2](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=2&annotation=YGXIRTPW)) #5fb236
- “} .” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=E5H6AUJM)) #ff6666
- “right side introduces adversarial training to mitigate the model overfitting of the training process” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=PB5W9LC5)) #5fb236
- “extracted from issues and commits to improving the model’s comprehension.” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=NB6V5APH)) #5fb236
- “I Di represents the issue description” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=S62LIPDK)) #5fb236
- “"p” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=SBU8KR9E)) #ff6666
- “entropy loss function to train the model:” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=PJIDV3BN)) #5fb236
- “Different templates may have distinct advantages and be suited to different dataset scenarios. Existing research lacks prior knowledge of which prompt template is most effective.” ([“ESEM24_paper_220.pdf”, p. 3](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=3&annotation=PTLW6DP6)) #ffd400
*SO far the paper is not easy to ready. I would have preferred to read some explanatory examples, on the quest of finding a right tradeoff between completeness and readability/understandability of the proposed approach.*
- “six open-source projects” ([“ESEM24_paper_220.pdf”, p. 4](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=4&annotation=FGX9Y735)) #5fb236
- “calculated means and standard deviation of each metri” ([“ESEM24_paper_220.pdf”, p. 4](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=4&annotation=XH4EP8UW)) #5fb236
- “a degree of stability and consistency in their performance across projects. I” ([“ESEM24_paper_220.pdf”, p. 4](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=4&annotation=D9XNQWIB)) #5fb236
- “RoBERTa, GPT-2, and BERT exhibit good stability. Notably, RoBERTa achieves the best average performance and the slightest fluctuations across all measures, indicating excellent generalisation capability.” ([“ESEM24_paper_220.pdf”, p. 5](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=5&annotation=4EPW29RV)) #5fb236
- “PromptLink will default to using RoBERTa for artifact representation and mask prediction.” ([“ESEM24_paper_220.pdf”, p. 5](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=5&annotation=GWLDNURN)) #5fb236
- “adversarial training is effective in preventing model overfitting” ([“ESEM24_paper_220.pdf”, p. 6](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=6&annotation=Q53U27LE)) #5fb236
- “template types and their applicability in different scenario” ([“ESEM24_paper_220.pdf”, p. 6](zotero://select/library/items/33T6XVEN)) ([pdf](zotero://open-pdf/library/items/2TVISAU2?page=6&annotation=GX9QHZFU)) #5fb236
- ### [[Comments]]
- Overview: The paper presents PromptLink, a multi-template prompt learning approach combined with adversarial training to address the problem of issue-commit link recovery.
The approach has been experimented on six open-source projects, revealing that PromptLink outperforms state-of-the-art methods concerning different metrics, including F1, Precision, Recall, MCC, AUC, and ACC.
- Comments: My main concerns about the paper are related to its presentation and clarity. The document is not always easy to read. It could benefit from more explanatory examples to enhance readability and understanding, especially for readers less familiar with the intricacies of prompt learning and adversarial training. Even the background and motivation could have been better presented. Explanatory examples should be presented to distill the limitations of existing approaches in real cases. It is necessary to find the right tradeoff between clarity of presentation and completeness and formality of the presented technical concepts.
- ### [[REVIEWS/Notes]]
- ### YELLOW CONCERNS
background-color:: yellow
- {{query (and [[ffd400]] [[ESEM24_paper_220]] )}}
collapsed:: true
- ### ❓️Questions
- {{query (and [[question]] [[ESEM24_paper_220]] )[[question]]}}
query-table:: true
query-properties:: [:block]