An Aspect-Based Emotion Analysis Approach on Wildfire-Related Geo-Social Media Data — A Case Study of the 2020 California Wildfires

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The reviewed paper is very well-written and meaningfully structured. My main comments concern the presented procedures' applicability, the research's replicability, and the transferability of results. These comments are related to the current (in)availability of data from the X network (formerly Twitter), or rather, differences in the use of this social media over time and space.

I believe the paper's authors will be able to respond to all my comments. In general, I recommend the paper for acceptance after major revisions. I hope my comments will help the authors improve the quality of the paper. Individual comments are listed below. 

Introduction

• The authors present the issue of emotion analysis here. I want to add that Grace (2020; https://doi.org/10.1016/j.dib.2020.105595) also addressed this issue. They define slightly different categories of emotions, although there is some overlap. I have personal experience that this approach has been applied to analysing social media during and after a crisis, not only X (Twitter).
• The authors state that their approach is a BERT-based model. Myint, Lo and Zhang (2024; https://doi.org/10.1016/j.ipm.2024.103695) also test and extend this type of model on similar data from social media.

Related works

• Studies that have tried to use social media for analysing crises, specifically fires, are also significantly earlier; for example, Laura Spinsanti and Ostermann (2013; https://doi.org/10.1016/j.apgeog.2013.05.005) or Kent and Capello (2013; https://doi.org/10.1080/15230406.2013.776727).

Methodology

• Section 3.1: Did the fact that it is located in the United States of America, which is relatively densely populated in the California area and probably has the highest number of Twitter users, also play a role in choosing the event under study? In 2020, there were also wildfires in Australia, which affected a much larger area (over 60 million acres). Still, the bush is significantly less populated, so Twitter is used significantly less there.
• Section 3.2: Why were a radius of 30 km and an interval of two weeks chosen when creating data subsets? I recommend adding an explanation and justification for these parameters' values.
• The authors used, among others, 3D visualisation to present the results, specifically an oblique view of extruded hexagons. They further expressed the ratio change through colour hue. I believe this representation method may be challenging for some users, so I would recommend adding a short section to the methodology that will explain the chosen methods of (cartographic) visualisation. For this purpose, see, for example, Shepherd (2008; https://doi.org/10.1002/9780470987643.ch10).

Results

• Second paragraph: This section could be moved to the methodology, to the section on cartographic visualisation proposed above.
• Figures 4, 5, as well as further 9, 10, and 11: Legend intervals are not unambiguous - they overlap. See, for example, Golebiowska, Korycka-Skorupa and Slomska-Przech (2021; https://gistbok-topics.ucgis.org/CV-04-011)
• My subjective opinion is that the quality of the interpretation of the results would benefit if the author team included someone with more profound local knowledge of the research area. However, I do not expect the authors to make any changes in this regard.

Discussion

• Authors should consider including answers to the following questions in the discussion:
  • What is the current (in 2025) applicability of data from X (formerly Twitter)?
  • Are the presented procedures and results applicable to other social media than just X (Twitter)?
  • Are the presented outcomes also applicable in regions other than the USA? How does this affect a given social media platform's market penetration or the multiple languages ​​used in other regions?
• Figure 11:
  • Legend for the "Anger" ratio is arranged in reverse order from the previous maps. I recommend unifying (and editing) the legends as much as possible. This is about the comparability of individual fire locations.
  • These 3D maps should also contain the north arrow for better orientation.

Conclusion

• Wouldn't it be possible to add a summarising paragraph about the follow-up research ("future work")?

References

• [31]: The citation of this source should be extended.
• [33]: The URL should be added here

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper presents a novel application of aspect-based emotion analysis (ABEA) to geo-social media data, focusing on Twitter posts during the 2020 California wildfires. The authors apply a fine-tuned BERT-based model (EmoGRACE) to identify emotional responses (e.g., anger, sadness, fear, happiness) linked to specific aspects (like “fire”, “evacuation”, etc.) in spatio-temporal context.

The study aims to reveal how emotional discourse evolves near wildfire perimeters over time, providing insights potentially valuable for emergency response and disaster management. It’s a very practical research topic. To further improve the quality of the manuscript, my comments and suggestions are as follows:

1. Model Performance and Validation:

The authors acknowledge EmoGRACE’s limited performance, especially in emotion classification. Is there any quantitative evaluation (e.g., accuracy, F1-score) on held-out wildfire-specific data to support the model's reliability for this task?

2. Model Generalizability:

The study focuses on only four wildfires (from an initial ten), limiting broader applicability. Could the authors discuss how well the approach might generalize to other disaster types or less-documented events?

3. About the Emotion Label Granularity:

The paper groups nuanced emotions under four broad classes. Could more detailed emotional categories (e.g., “relief”, “anxiety”, “frustration”) offer better insight for policy-making or disaster communication?

4. MAUP Handling:

The paper briefly addresses the Modifiable Areal Unit Problem. Still, further clarification is needed on whether the selected 5×5 km grid is optimal for both urban and rural zones, where tweet densities differ drastically.

5. Practical Impact:

What specific examples can the authors offer where this type of emotional mapping influenced (or could influence) real-time decision-making in emergency management?

Based on the above considerations, I would like to give you a major revision to further improve the quality of the manuscript.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The author analysis the emostion on wildfire, but the novelty is very limited in the ascpect of scientifical view. I have following suggestions for improving the article:

1. The drawback of exisiting emotion analysis approach should be introduced to hightlight the novelty. Meanwhile, the sepecifical issues related to widefire need to be discussed as well.
2. The author use existing EmoGRACE model for emotion analysis. The proposed approach should be improved to solve certain problem, but I can hardly find the improvement nor the problem.
3. The proposed method should be compared with other existing approaches to illustrate the performance
4. It seems to be more suitable for social science journals.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

All my suggestions have been responded to in the text of the paper and explained in the provided cover letter. The clarity of the paper has increased.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper has been well refined. No further question/suggestion.

Reviewer 3 Report

Comments and Suggestions for Authors

I think the paper is revised well and can be published now.