C-STEER: A Dynamic Sentiment-Aware Framework for Fake News Detection with Lifecycle Emotional Evolution

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

• Clarify the empirical measurement of theoretical constructs to strengthen the link between theory and computation.

• Acknowledge that Weibo16 and Twitter16 are dated; suggest validation on newer or multilingual misinformation datasets.

• Discuss computational cost and scalability compared to recent large language models (LLMs).

• Include a short section on ethical implications, especially regarding privacy and bias in emotion analysis.

• Simplify overly technical portions (Sections 3.3–3.6) or move detailed equations to supplementary material to improve readability.

• Quantify improvements in the Abstract

• Improve figure readability and reference formatting (add spaces before citations).

• Add a short “Limitations and Future Work” paragraph in the conclusion.

Author Response

We would like to express our sincere gratitude to you for the insightful and constructive comments. These suggestions have significantly helped us improve the theoretical grounding, readability, and rigor of our manuscript. We have carefully addressed all points raised.

Comment 1: Clarify the empirical measurement of theoretical constructs to strengthen the link between theory and computation.

Response 1: Thank you for this valuable suggestion. We agree that explicitly linking the abstract theoretical constructs to our computational modules is essential for the paper's logic. Therefore, we have revised Section 3.4.1 to clearly map the concepts from Uses and Gratifications Theory and Emotional Contagion Theory to our specific vector dimensions. Specifically, we clarified how the user's "need for venting" is operationalized as the Negative-Valence Ratio, and how "collective volatility" is measured via Emotion Standard Deviation. This ensures the "Theory-Module-Feature" alignment is transparent.

[Changes made in the manuscript]:

• Revised the introductory paragraph of Section 3.4.1 to include the mapping between theory and metrics.

 

Comment 2: Acknowledge that Weibo16 and Twitter16 are dated; suggest validation on newer or multilingual misinformation datasets.

Response 2: We strongly agree that analyzing up-to-date data is crucial. To address this, we have replaced the dated Weibo16 dataset with the newer Weibo21 dataset in our experiments. Weibo21 covers more recent news events with multimodal content, presenting a harder challenge. Despite this update, Twitter16 remains a widely used benchmark for cross-lingual comparison. Therefore, we have retained Twitter16 but added a explicit statement in the Limitations (Section 5) acknowledging the age gap between datasets and the rapid evolution of social media content forms.

[Changes made in the manuscript]:

• Experiments updated with Weibo21 throughout Section 4;

• Limitations discussion added in Section 5.

 

Comment 3: Discuss computational cost and scalability compared to recent large language models (LLMs).

Response 3: We have added a new subsection, Section 4.8 (Computational Efficiency and Scalability). We report that C-STEER operates with an average inference time of approx. 100 ms per event (on Weibo21), which is significantly more efficient than resource-heavy LLMs, making it suitable for real-time detection.

[Changes made in the manuscript]:

• Added Section 4.8.Computational Efficiency and Scalability

 

Comment 4: Include a short section on ethical implications...

Response 4: We have added Section 4.10 (Ethical Considerations). We clarify that all data (Weibo21/Twitter16) are anonymized public benchmarks. We also discuss potential algorithmic biases in emotion detection and advocate for human-in-the-loop auditing.

[Changes made in the manuscript]:

• Added Section 4.10 Ethical Considerations .

 

Comment 5: Simplify overly technical portions (Sections 3.3–3.6) or move detailed equations to supplementary material to improve readability.

Response 5: Thank you for pointing out the complexity. We have carefully reviewed Sections 3.3–3.6. To improve readability while preserving the logical completeness of the methodology in the main text, we have simplified the textual descriptions surrounding the equations. We removed redundant definitions and streamlined the explanation of standard mechanisms (e.g., standard LSTM gates) to focus the reader's attention on our novel contributions (adaptive segmentation and attention). The revised text is now more concise and approachable.

[Changes made in the manuscript]:

• Streamlined text and removed redundant descriptions in Sections 3.3–3.6.

 

Comment 6: Quantify improvements in the Abstract.

Response 6: We agree that specific numbers make the abstract more compelling. We have updated the Abstract to reflect the new experimental results. We now explicitly state that C-STEER achieves F1-macro scores of 91.6% and 90.1% on Weibo21 and Twitter16, respectively, highlighting the quantitative margin of improvement over baselines.

[Changes made in the manuscript]:

• Updated the Abstract with new F1 scores based on Weibo21 and Twitter16 experiments.

 

Comment 7: Improve figure readability and reference formatting (add spaces before citations).

Response 7: We apologize for the formatting oversight. We have replaced all Figures with high-resolution versions to ensure legibility. Furthermore, we have proofread the manuscript and corrected the citation formatting, ensuring a space is inserted before every citation bracket (e.g., changing "text[1]" to "text [1]").

[Changes made in the manuscript]:

• Updated all Figures with high-resolution versions.

• Corrected citation spacing globally.

 

Comment 8: Add a short “Limitations and Future Work” paragraph in the conclusion.

Response 8: Thank you for this suggestion. We have restructured the Conclusion. We added "Limitations and Future Work" paragraph In section5, we discuss the limitations regarding the heuristic nature of our segmentation and the age of the Twitter16 dataset (despite the update to Weibo21), and outline future plans for learnable segmentation and multimodal integration.

[Changes made in the manuscript]:

• Added a short paragraph in Section 5 to explicitly address limitations and future directions.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript entitled C-STEER: A Dynamic Sentiment-Aware Framework for Fake News Detection with Lifecycle Emotional Evolution proposes a disinformation detection model grounded in communication theory. Its core concept is to divide the dissemination process of a news item into three stages (initiation, burst, and decay) and to model the collective emotional evolution within each phase. This information is combined with textual and structural features through a deep learning architecture. The authors report improved performance on two datasets (Weibo16 and Twitter16) and conclude that emotional dynamics are a predictive factor for identifying false news.

The study represents a solid effort to bridge communication theory and computational modeling and relies on an extensive experimental evaluation. However, its contribution depends almost entirely on the validity of the temporal segmentation model, which at the same time constitutes the weakest point of the work. The following elements summarize the main findings:

1. The division into three phases is adopted heuristically and uniformly, without empirical justification to demonstrate its general applicability. Interactions on social networks do not always follow a unimodal pattern, hence, fixed segmentation may induce misalignment errors and loss of temporal coherence.
2. The model assumes that emotions evolve in a regular manner (from curiosity to anger and then to skepticism), whereas there are multiple contexts (prolonged events, real crises, or institutional campaigns) that do not follow such transitions. In these cases, segmentation may degrade performance or produce false positives.
3. The manuscript itself acknowledges the heuristic nature of the temporal boundaries and mentions the need for a learnable segmentation procedure. This admission confirms that the most innovative part of the system still lacks sufficient theoretical and methodological grounding.
4. The model’s external validity is limited by the use of relatively small and outdated datasets. Generalization to other languages, cultures, or platforms remains untested.
5. Conceptually, the work occasionally conflates correlation with causation: the fact that emotional diffusion accompanies misinformation does not necessarily mean it is a determining indicator of falsity. This distinction should be treated with greater theoretical precision.

In consequence, this reviewer considers that the potential contribution of the manuscript lies in the formalization of the stage-based verification model. To make this contribution convincing, it will be necessary to reassess and strengthen its theoretical foundations, explore more flexible segmentation methods, and broaden the empirical evidence supporting its validity across diverse contexts.

Major revisions are recommended before the manuscript can be considered for possible publication.

Author Response

We would like to express our sincere gratitude to you for the critical and comprehensive assessment. We appreciate the recognition of our effort to bridge communication theory with computational modeling.

We took the major concerns regarding the validity of segmentation, dataset outdatedness, and the distinction between correlation and causation very seriously. In response, we have performed major revisions, most notably replacing the entire Weibo16 dataset with the newer Weibo21 dataset to demonstrate robustness on modern data, and adding a dedicated rationale section to justify the segmentation model with empirical evidence.

Comment 1: The division into three phases is adopted heuristically and uniformly... Interactions on social networks do not always follow a unimodal pattern, hence, fixed segmentation may induce misalignment errors.

Response 1: We appreciate this insightful comment regarding the complexity of propagation patterns. We have revised the manuscript to clarify that our work does not assume a rigid or strictly unimodal structure.

1. Empirical Justification: While social propagation is complex, extensive prior research confirms that misinformation, specifically, consistently follows a "burst -> main peak -> decay" dominant dynamic. As we now explicitly discuss in the newly added "Rationale and Scope" paragraph in Section 3.3 (Page 6, Lines 354-356), large-scale studies by Zhao et al. [35] and Pfeffer et al. [36] demonstrate that fake news typically exhibits early explosive growth forming a clearly identifiable dominant peak. Our model is designed to target this dominant trajectory.

2. Adaptive Implementation: Crucially, our method is data-adaptive, not fixed. As detailed in Section 3.3.1 (Page 9, Eq. 1), we employ an Adaptive Sliding Window that adjusts window size based on real-time propagation velocity (using higher resolution during bursts). Furthermore, our peak detection algorithm anchors the segmentation to the global dominant peak ($P_{main}$). Even in multi-peak scenarios, this ensures the model consistently identifies the most discriminative "Burst" phase centered on maximum volatility, maintaining temporal coherence.

[Changes made in the manuscript]:

• Page 6, Lines 354-356: Added a "Rationale and Scope" paragraph explicitly citing empirical studies (Zhao et al., Pfeffer et al.) to justify the main-peak-centered segmentation.

• Page 9, Figure 3: Updated the flowchart to visually emphasize the adaptive loop (speed sensing -> window adjustment).

Comment 2: The model assumes that emotions evolve in a regular manner (from curiosity to anger to skepticism)... whereas there are multiple contexts that do not follow such transitions.

Response 2: This is a fair point. We acknowledge that individual news events (e.g., prolonged crises) may vary.

However, our core argument is probabilistic: fake news relies on this specific emotional manipulation script to go viral, whereas real news often defies it. Therefore, capturing this specific "initiation-burst-decay" emotional arc serves as a discriminative filter.

The Ablation Study (Table 7, Page 20) on the new Weibo21 dataset confirms this: removing the LifeCycle-Emotion module causes a statistically significant performance drop (-1.8% F1-macro). This indicates that despite individual variations, the aggregate statistical regularity captured by our segmentation provides non-redundant predictive power that general-purpose models miss.

[Changes made in the manuscript]:

• Page 2, Introduction: Clarified that the model captures "stage-wise emotion dynamics" as cues for "disentangling different propagation drivers"⁴.

• Page 20, Table 7: Provided ablation results on Weibo21 proving the module's contribution.

Comment 3: The admission of the heuristic nature confirms that the most innovative part lacks sufficient theoretical and methodological grounding.

Response 3: We understand the concern. However, we view the parameterization not as arbitrary "heuristics," but as "theory-guided operationalization."

Every threshold (e.g., the pre-peak window) is derived from the quantitative criteria of Diffusion of Innovations Theory regarding "Take-off" and "Saturation" points. To strengthen the methodological grounding, we have revised Section 3.4.1 (Page 11-12) to explicitly map abstract theoretical constructs to our computable vectors. For instance, we link the "Need for Venting" (Uses and Gratifications) directly to the Negative-Valence Ratio feature, and "Collective Volatility" to Emotion Standard Deviation. This transforms the "heuristic" into a transparent, theory-grounded White-box approximation.

[Changes made in the manuscript]:

• Page 11-12, Section 3.4.1: Added explicit mapping between Uses and Gratifications concepts and specific feature dimensions to strengthen theoretical grounding.

Comment 4: The model’s external validity is limited by the use of relatively small and outdated datasets.

Response 4: We have critically accepted this comment. We agree that relying on Weibo16 limits validity regarding modern social media.

Action Taken: We have replaced the Weibo16 dataset with the newer Weibo21 dataset throughout the entire paper. Weibo21 (4,488 fake, 4,640 real) represents a significantly larger and more modern benchmark.

We re-trained C-STEER and all 11 baselines on Weibo21. As shown in Table 5 (Page 18), C-STEER achieves an F1-macro of 91.6%, significantly outperforming the state-of-the-art TDEI model (90.3%). This proves that our lifecycle-based framework remains highly effective on modern, large-scale data. We retained Twitter16 to ensure cross-lingual comparability but acknowledged its age in the Limitations.

[Changes made in the manuscript]:

• Page 17, Section 4.1.1: Updated dataset description to Weibo21.

• Pages 18-24: Updated Table 5, Table 7, Table 8, Table 9, and all corresponding figures (Fig 5-9) with new experimental results from Weibo21.

Comment 5: Conceptually, the work occasionally conflates correlation with causation: the fact that emotional diffusion accompanies misinformation does not necessarily mean it is a determining indicator of falsity.

Response 5: This is a crucial distinction. We have refined our theoretical framing to emphasize causality grounded in user intent.

Drawing on Uses and Gratifications Theory (Section 3.4), we posit that high-arousal emotion is a causal driver, not just a byproduct. Users actively utilize emotions (e.g., venting anger) to satisfy psychological needs, which accelerates propagation. Fake news is engineered specifically to exploit this mechanism. Therefore, our "LifeCycle-Emotion" features are designed to detect the traces of this specific manipulation, distinguishing the engineered virality of fake news from the organic spread of factual events.

[Changes made in the manuscript]:

• Page 11, Section 3.4: Expanded the explanation of Uses and Gratifications to frame emotional expression as a need-driven, causal driver of propagation.

Comment 6: In consequence, this reviewer considers that the potential contribution... lies in the formalization of the stage-based verification model. To make this contribution convincing... reassess and strengthen foundations... broaden empirical evidence.

Response 6: We believe the major revisions undertaken—specifically replacing the dataset with Weibo21 (broadening empirical evidence), adding the Rationale and Scope section (strengthening foundations), and clarifying the adaptive segmentation (addressing flexibility)—have directly addressed the reviewer's core requirements. The updated results demonstrate that the stage-based model is robust across languages (Chinese/English) and time periods (2016/2021).

[Changes made in the manuscript]:

• Page 25, Section 5: Restructured the Conclusion to synthesize these contributions and findings cohesively.

Reviewer 3 Report

Comments and Suggestions for Authors

A brief summary 

The paper presents a novel framework grounded in communication theory for fake-news detection.

The aim of this paper is to is to increase effectiveness of fake new detections.

The main contribution is new framework for  fake news detection. The framework is based on the theory of the diffusion of innovations, which divides the spreading process into three stages: start-up, burst, and decay. Utilizing the theories of Emotional Contagion and Uses and Gratifications, the framework establishes edge weights and identifies emotion features that are unique to each stage. After getting attention, a BiLSTM learns how these emotional signals change over time. Finally,  all this signals are used  for classification. 

Broad comments 

1) It is unclear which metric is used to calculate the Life-Cycle Segmentation of News. Whether it is the number of messages or some other quantity characterizing the communication.

2) An illustration of calculation on the chart of Life-Cycle Segmentation of News, presented on the chart, would be very helpful.

Specific comments:

l. 231 - charts are barely visible. Presenting them one by one would improve the situation.

l. 273 - integer division M//20 is unclear. Can it be defined with division (/) and floor or ceiling operation,

Author Response

We sincerely thank you for the positive summary and the constructive comments regarding the clarity of our definitions and visualizations. These suggestions have helped us significantly improve the readability and rigor of the manuscript. We have addressed the comments point-by-point below.

 

Broad Comment 1: It is unclear which metric is used to calculate the Life-Cycle Segmentation of News. Whether it is the number of messages or some other quantity characterizing the communication.

Response 1: Thank you for pointing out this ambiguity.

We have explicitly defined the metric in Section 3.3.1 to ensure clarity. We specified that the metric $\Delta c$ used for calculating propagation speed and subsequent segmentation refers to the total count of user interactions (summing retweets, comments, and likes) within the sliding window. This aggregate count captures the overall volume of user engagement driving the diffusion process.

[Changes made in the manuscript]:

• Page 9, Lines 266-268: "Let ∆c be the total count of user interactions (summing retweets, comments, and likes) within the sliding window..."

 

Broad Comment 2: An illustration of calculation on the chart of Life-Cycle Segmentation of News, presented on the chart, would be very helpful.

Response 2: We agree that a visual illustration aids in understanding the segmentation logic.

In Figure 3 (Page 9), we present a schematic flowchart of the Adaptive Life-Cycle Segmentation algorithm. This diagram visually maps out the calculation flow: from "Compute propagation velocity" using dynamic windows, to "Peak detection" ($P_{main}$), and finally to the "Phase split" based on the detected boundaries. This visual aid, combined with the detailed criteria in Table 2 and the equations in Section 3.3.2, provides a comprehensive illustration of how the segmentation is calculated.

[Changes made in the manuscript]:

• Page 9, Figure 3: Updated the flowchart to clearly illustrate the calculation steps from velocity computation to phase splitting.

• Page 9-10, Section 3.3: Refined the accompanying text to better explain the workflow shown in the figure.

 

Specific Comment 1 (l. 231): charts are barely visible. Presenting them one by one would improve the situation.

Response 3: We apologize for the poor visibility in the previous version. We have completely reformatted Figure 2. Instead of squeezing the subplots horizontally, we have now arranged them vertically (stacked as Figure 2a, 2b, and 2c) and significantly increased the resolution and font size. This ensures that the distinct trends for Awareness, Interpersonal Share, and Sentiment Intensity are clearly visible and easy to compare.

[Changes made in the manuscript]:

• Pages 7-8, Figure 2: Re-plotted and vertically stacked the subplots (Fig 2a, 2b, 2c) for maximum readability.

 

Specific Comment 2 (l. 273): integer division M//20 is unclear. Can it be defined with division (/) and floor or ceiling operation.

Response 4: Thank you for the correction regarding mathematical notation.

We have replaced the Python-style integer division code (//) with the standard mathematical floor function notation ⌊·⌋.

[Changes made in the manuscript]:

• Page 10, Lines 280-282: Revised to: "Width. Set to ⌊M/ 20⌋ (where ⌊·⌋ denotes the floor function)... Distance. Set to ⌊M/ 10⌋..."

Reviewer 4 Report

Comments and Suggestions for Authors

The manuscript presents an innovative framework that integrates communication theory with deep learning to model the emotional dynamics of fake news dissemination. The topic is timely and relevant to both computational social science and affective computing. Overall, the paper demonstrates methodological rigor, sound experimental validation, and a well-structured theoretical foundation.
1- The authors should avoid using bold font within the Related Work section, as it is not aligned with scientific writing conventions. 
2- Abbreviations should be written in full at their first occurrence. For example, C-STEER and LSTM should be introduced with their complete forms in the abstract.
3- Ensure that all abbreviations are fully defined. For instance, GCN and GAT should be written as Graph Convolutional Network (GCN) and Graph Attention Network (GAT) when first mentioned.
4- The Related Work section would be more informative if it included the reported performance metrics (e.g., accuracy, recall, F1-score) of cited studies. 
5- The figures throughout the manuscript are unclear. Please enhance their resolution, and adjust labeling.
6- Remove the redundant line “Summary. This section presents a fake-news” in Section 3, as it does not contribute meaningfully to the narrative.
7- The Model Details section is difficult to follow and should be reorganized for clarity. 
8- The section titled Overview of the C-STEER Architecture includes two figures. It should be clarified which one represents the proposed model.
9- Avoid subsections or sub-subsections that contain only one paragraph, such as “4.1.3 Implementation Details.” Either merge these into broader sections or expand them with more substantive content.
10- The term “ifeCycleEmotion” is unclear and should be defined precisely when first introduced.
11- The Conclusion section should be condensed into a single cohesive paragraph summarizing the key contributions, findings, and potential future directions. Subsections are unnecessary for this part of the paper.

Comments on the Quality of English Language

Review the manuscript for linguistic clarity. Simplifying complex sentences and avoiding redundancy will improve overall readability and scientific precision.

Author Response

We sincerely thank you for the encouraging feedback and the meticulous attention to detail regarding scientific writing conventions and manuscript structure. We found the comments extremely helpful in refining the presentation of our work. We have carefully addressed all formatting and structural suggestions.

Comment 1: The authors should avoid using bold font within the Related Work section, as it is not aligned with scientific writing conventions.

Response 1: Thank you for correcting our formatting. We have removed the bold formatting in Section 2 (Related Work). We now use italics (e.g., Traditional machine learning, Graph-based modeling) to denote sub-themes, strictly adhering to standard scientific writing conventions.

[Changes made in the manuscript]:

• Page 2-3, Section 2: Removed bold fonts and applied standard formatting.

Comment 2: Abbreviations should be written in full at their first occurrence. For example, C-STEER and LSTM should be introduced with their complete forms in the abstract.

Response 2: We have reviewed the Abstract and ensured all abbreviations are fully defined upon first use. We explicitly defined "C-STEER" as (Cycle-aware Sentiment-Temporal Emotion Evolution) and "BiLSTM" as (Bidirectional Long Short-Term Memory) in the Abstract.

[Changes made in the manuscript]:

• Page 1, Abstract: Added full definitions for C-STEER and BiLSTM.

Comment 3: Ensure that all abbreviations are fully defined. For instance, GCN and GAT should be written as Graph Convolutional Network (GCN) and Graph Attention Network (GAT) when first mentioned.

Response 3: We have scanned the manuscript and corrected all undefined abbreviations. Specifically, in Section 2 (Related Work), we now provide the full terms for Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) at their first occurrence.

[Changes made in the manuscript]:

• Page 3, Section 2: Added full definitions for GCN and GAT.

Comment 4: The Related Work section would be more informative if it included the reported performance metrics (e.g., accuracy, recall, F1-score) of cited studies.

Response 4: We appreciate this suggestion to add quantitative context. However, after careful consideration, we decided not to include specific performance metrics in the Related Work section. The primary reason is that the cited studies (e.g., SSE-BERT, TDEI, SA-HyperGAT) often use different datasets (e.g., PolitiFact vs. Weibo vs. Twitter), different splitting strategies (temporal vs. random), and different evaluation protocols. Listing their reported numbers side-by-side in Section 2 could be misleading due to the lack of a common baseline. Instead, we provide a rigourous, direct comparison in Section 4 (Experiments), where we reproduce these baselines on the same datasets (Weibo21 and Twitter16) under identical experimental settings. We believe this offers a fairer and more scientifically valid assessment of relative performance.

Comment 5: The figures throughout the manuscript are unclear. Please enhance their resolution, and adjust labeling.

Response 5: We apologize for the low resolution in the previous version. We have replaced all figures (Figures 1-9) with high-resolution vector or high-DPI images. We also adjusted the font sizes and layouts (e.g., vertically stacking Figure 2) to ensure that all legends, axes, and labels are clearly legible.

[Changes made in the manuscript]:

• Updated all Figures in the manuscript with high-quality versions.

Comment 6: Remove the redundant line “Summary. This section presents a fake-news” in Section 3, as it does not contribute meaningfully to the narrative.

Response 6: We agree. We have removed the redundant summary paragraph at the end of Section 3.6 to improve the flow and conciseness of the narrative.

[Changes made in the manuscript]:

• Page 16, Section 3.6: Removed the redundant summary text.

Comment 7: The Model Details section is difficult to follow and should be reorganized for clarity.

Response 7: Thank you for this feedback. To improve clarity, we have added a "roadmap" paragraph at the beginning of Section 3.1 (Overall Framework). This paragraph explicitly guides the reader through the logical flow of the subsequent subsections: from Graph Construction (§3.2), Life-Cycle Segmentation (§3.3), Emotion Extraction (§3.4), to Text Encoding (§3.5) and Fusion (§3.6). We also streamlined the descriptions in Sections 3.3–3.6 to make the technical details more accessible.

[Changes made in the manuscript]:

• Page 4, Section 3.1: Added a paragraph outlining the structure of Section 3.

Comment 8: The section titled Overview of the C-STEER Architecture includes two figures. It should be clarified which one represents the proposed model.

Response 8: We have clarified the distinct roles of the two figures in their captions. Figure 1 represents the "Overview of the C-STEER architecture," showing the entire multi-modal pipeline. Figure 4 represents the "Internal structure of the STEER-Encoder," which is a specific sub-module for encoding graph and emotion sequences. The captions now clearly reflect this hierarchy.

[Changes made in the manuscript]:

• Page 4, Figure 1 Caption: Explicitly labeled as "Overview of the C-STEER architecture."

• Page 14, Figure 4 Caption: Explicitly labeled as "Internal structure of the STEER-Encoder..."

Comment 9: Avoid subsections or sub-subsections that contain only one paragraph, such as “4.1.3 Implementation Details.” Either merge these into broader sections or expand them with more substantive content.

Response 9: We have followed this advice. We merged the content of the original "Implementation Details" into a broader section titled 4.1.2 Experimental Settings. This section now coherently covers metrics, statistical tests, hyperparameters, and environment setup in a unified manner.

[Changes made in the manuscript]:

• Page 17, Section 4.1.2: Merged implementation details into "Experimental Settings".

Comment 10: The term “ifeCycleEmotion” is unclear and should be defined precisely when first introduced.

Response 10: Thank you for spotting this ambiguity (and the likely typo). We have formally defined the term "LifeCycle-Emotion" in Section 3.4.1 (where the feature vector is constructed). We explicitly state: "For clarity in the subsequent experimental analysis, we refer to this specific 13-dimensional feature set capturing stage-wise dynamics as LifeCycle-Emotion." This ensures the term is precisely defined before it appears in the Ablation Study.

[Changes made in the manuscript]:

• Page 12, Section 3.4.1: Added a formal definition sentence for "LifeCycle-Emotion".

Comment 11: The Conclusion section should be condensed into a single cohesive paragraph summarizing the key contributions, findings, and potential future directions. Subsections are unnecessary for this part of the paper.

Response 11: We have restructured the Conclusion. We removed the fragmented subsections (e.g., "Findings," "Contributions") and synthesized the content into a single, cohesive narrative in Section 5. We retained only a separate, short subsection for "Limitations and Future Work" (Section 5.3) to clearly highlight future directions, as suggested by other reviewers, but the main conclusion is now unified.

[Changes made in the manuscript]:

• Page 25, Section 5: Condensed the conclusion into a unified narrative.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have adequately addressed the comments raised in the previous round. Revisions clarify the methodological assumptions, theoretical grounding, and update the experimental validation (more appropriate dataset). The concerns regarding segmentation, emotional dynamics, and empirical support have been resolved.

I recommend acceptance of the manuscript in its current form.