All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
The reviewer is satisfied with the recent changes proposed by the authors and therefore I can recommend this article for acceptance.
[# PeerJ Staff Note - this decision was reviewed and approved by Mike Climstein, a PeerJ Section Editor covering this Section #]
The authors have completely addressed all my comments, and I have no further concerns. Therefore, I recommend accepting the paper.
-
-
Please clarify the contribution of this work in light of existing studies highlighted by Reviewer 4.
**PeerJ Staff Note:** Please ensure that all review and editorial comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
Some minor grammatical errors exists
Although the paper explains why BERT and CNN are used together, it lacks a deeper justification for specific architecture choices (e.g., why Conv1D filters = 128, or why kernel size = 5)
Improve Figure 1 and Figure 2 with clearer annotations and labels.
The OnSIDe BERT-CNN outperforms the baseline models on Reddit (F1 = 92.88%) and performs robustly on unseen Twitter data (F1 = 83.81%).
The manuscript is generally well written and addresses a highly relevant and sensitive topic—detecting suicidal ideation from social media posts using a BERT-CNN hybrid model. The use of professional English is mostly consistent, although a few sections (particularly the methodology and results discussion) would benefit from clearer and more concise phrasing.
The introduction effectively explains the context and importance of early suicide detection in online environments. The motivation behind the proposed method is clear, and the related work is adequately cited. However, some recent and closely related work in the domain of social media-based mental health detection using deep learning could be referenced to further enrich the literature review.
One such study that would support the context and strengthen the relevance is:
Deep learning-based detection of depression and suicidal tendencies in social media data with feature selection. This work highlights similar applications of deep learning in social mental health detection, with an emphasis on feature-driven model design.
The overall structure of the manuscript aligns well with publication standards. Figures and tables are informative but could be better integrated with the main discussion for clearer interpretation.
The study fits well within the journal’s scope, and the research question is both timely and clearly defined. The dataset used is appropriate, and the use of BERT for transfer learning is justified. The integration of CNN to enhance local feature learning is also reasonable and complements the BERT embeddings. That said, the experimental setup could be described in more detail. For example, it would help to include more specifics on how the text was preprocessed before feeding into BERT, the exact hyperparameters used, and how the final BERT-CNN architecture was tuned. This information is important for reproducibility. The authors do mention that the code is shared via GitHub, which is appreciated. However, a brief description of the repository structure and instructions for running the model would enhance transparency. It would also be beneficial to include a short ablation study or component analysis to show the contribution of each part of the model (e.g., BERT-only vs. BERT+CNN).
The results reported are strong, with high accuracy on both Reddit and Twitter datasets. However, the manuscript would benefit from a brief error analysis or discussion of model limitations, such as challenges with short posts, sarcasm, or ambiguous language. While the findings align with the objectives stated in the introduction, the conclusion could better reflect the limitations and potential areas for improvement—such as generalizability across platforms or cultural variations in language use. Including a discussion on how this model could be deployed in real-world applications would strengthen the practical impact.
This paper tackles an important issue with a thoughtful technical approach. With some additional clarification around the experimental setup, a more detailed error or ablation analysis, and minor improvements in writing clarity, the manuscript would be a valuable contribution to the field.
Again, I encourage the authors to cite their related work:
Deep learning-based detection of depression and suicidal tendencies in social media data with feature selection, which reinforces the importance and potential of deep learning in psychological risk detection.
The manuscript is generally good and structured, but the contribution relative to prior hybrid models (such as BERT + CNN) needs clearer positioning. While ethical considerations surrounding suicide detection are only briefly touched on, this should be expanded to address broader societal risks. In addition, the figures are useful, but could benefit from improved readability and explanation, e.g., Figure 3 word clouds are somewhat crowded. I recommend explicitly highlighting novelty compared to existing work while emphasising the research impact in the introduction section. Also, adding a brief ethics discussion and societal impact in the conclusion will be helpful.
The experimental design is good, and the methods are reproducible, but the Twitter dataset’s preprocessing is less thoroughly described compared to the Reddit data. A stronger baseline, like RoBERTa, was discussed but not included. To strengthen the study, the authors should detail Twitter data cleaning steps, add a brief discussion on the absence of stronger baselines, and consider qualitative analysis (e.g., examples of misclassified cases) to complement quantitative metrics. You may also acknowledge in the methods section that stronger transformer baselines exist, even if not directly tested.
Findings are genaerally supported by the results, but broader validity could be enhanced with more detailed discussion on dataset limitations (e.g., demographic bias) and the challenges of generalising to non-English languages. A future work paragraph discussing multilingual extensions and a clearer explanation of performance trade-offs (e.g., between OnSIDe and RoBERTa) would help situate the findings more robustly.
The study is important and technically sound, with good reproducibility practices. Minor grammatical corrections and clearer emphasis on the practical benefits of the OnSIDe model (e.g., computational efficiency, application scenarios) would further strengthen the presentation. Overall, the manuscript is close to publishable with minor revisions focused on framing, impact, and clarity
The introduction is clear and well-structured, but it should more explicitly highlight the paper’s contributions, particularly in the context of suicidality detection. The authors also need to justify the choices made in their modeling approach to strengthen the relevance and impact of the work.
The paper proposes a BERT-CNN model for detecting suicidality in social media posts; however, the approach lacks novelty as similar architectures have been previously explored. Here some of the studies that uses Bert with CNN
Gorai, J., Shaw, D.K. A BERT-encoded ensembled CNN model for suicide risk identification in social media posts. Neural Comput & Applic 36, 10955–10970 (2024). https://doi.org/10.1007/s00521-024-09642-w
E. Lin, J. Sun, H. Chen and M. H. Mahoor, "Data Quality Matters: Suicide Intention Detection on Social Media Posts Using RoBERTa-CNN," 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Orlando, FL, USA, 2024, pp. 1-5, doi: 10.1109/EMBC53108.2024.10782647.
The authors use publicly available dataset but didn't compare the proposed method with latest approches in the field .
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.