Graph-Based Semi-Supervised Deep Learning for Indonesian Aspect-Based Sentiment Analysis
Round 1
Reviewer 1 Report
Comments on Graph-based Semi-supervised Deep Learning for Indonesian Aspect-Based Sentiment Analysis
I carefully evaluated the paper entitled Graph-based Semi-supervised Deep Learning for Indonesian Aspect-Based Sentiment Analysis. A range of problems can be found in this manuscript. Detailed comments are as follows:
1. Proofreading is recommended for this manuscript.
2. Line 203. Indentation is not recommended since ‘where’ does not begin with another sentence.
3. ‘a convolutional network’ refers to GCN which makes it confusing what CNN refers to since both GCN and CNN are used.
4. What are the main differences between Figure 5 and Figure 7? How to obtain the embeddings and how is adjacent matrix processed in a 1-dimension convolutional operation (Figure 5)? What do the node features consist of?
5. Table 6 and Table 7 should give results in consistent decimal places.
6. There is a peak between epoch 0~5, expatiate it.
7. Line 298-313. The figures could be improved, especially for the fonts.
8. Line 324-327. Why all the metrics in different methods are the same？GRN has only 0.25% improvement compared to GCN which may be due to randomness during model training. Averaging the repeated results is recommended.
9. Line 370. Total aspects.
10. Listing references in alphabetical order is recommended.
Author Response
Dear Reviewer,
We appreciate your precious time and effort in reviewing our manuscript and providing valuable comments. Your valuable and insightful advice has led to possible improvements in our manuscript. We have carefully considered and tried our best to revise all comments brought up by you as a reviewer. We hope the careful revision meets your high standards.
Below we provide the point-by-point responses. The original comments are provided in black color, whereas our answers are written in red.
We have also made corrections and refined the quality of the manuscript; these are also highlighted in yellow.
Sincerely,
Ahmad Abdul Chamid
Point 1: Proofreading is recommended for this manuscript.
Response 1: We thank you for your recommendation. We are considering using the proofreading provided by MDPI.
Point 2: Line 203. Indentation is not recommended since ‘where’ does not begin with another sentence.
Response 2: We thank you for your recommendation. We omit the word 'where' and write the equations explanation directly.
Point 3: ‘a convolutional network’ refers to GCN which makes it confusing what CNN refers to since both GCN and CNN are used.
Response 3: We also appreciate this comment. We use 'graph convolutional network' refers to (GCN) to detect aspects, and we also use convolutional neural network refers to (CNN) for sentiment classification. We have adjusted the manuscript.
Point 4: What are the main differences between Figure 5 and Figure 7? How to obtain the embeddings and how is adjacent matrix processed in a 1-dimension convolutional operation (Figure 5)? What do the node features consist of?
Response 4: We also thank you for your valuable comment. We get embedding using a word embedding tensorflow with https://www.tensorflow.org/text/guide/word_embeddings
We perform adjacent matrix processes for convolution using
https://www.tensorflow.org/api_docs/python/tf/keras/layers/Conv1D
The GCN architecture (Figure 5) uses the input layer = graph (we extract the text data into a graph using https://stanfordnlp.github.io/stanza/), while the CNN architecture (Figure 7) uses the input layer = text. The node features consist of dependency parser results such as
https://nlp.stanford.edu/software/lex-parser.shtml
Point 5: Table 6 and Table 7 should give results in consistent decimal places.
Response 5: We appreciate your comments to improve our manuscript. We have adjusted the manuscript to your suggestions.
Point 6: There is a peak between epoch 0~5, expatiate it.
Response 6: We thank you for your insightful suggestion. The effect of regulation and dropout makes accuracy out of the minimum locale. Then it dropped again because it entered the global area and returned to crawling up and being stable.
Point 7: Line 298-313. The figures could be improved, especially for the fonts.
Response 7: We appreciate your comments to improve our manuscript. We have adjusted the manuscript to your suggestions.
Point 8: Line 324-327. Why all the metrics in different methods are the same？GRN has only 0.25% improvement compared to GCN which may be due to randomness during model training. Averaging the repeated results is recommended.
Response 8: We thank you for your insightful suggestion. The same metrics are used because they are standard deep learning classifications. The 0.25 increase was obtained several times the process was repeated. The results are consistently better.
Point 9: Line 370. Total aspects.
Response 9: We appreciate your comments to improve our manuscript. We have adjusted the manuscript to your suggestions.
Point 10: Listing references in alphabetical order is recommended.
Response 10: We appreciate your comments to improve our manuscript. We use references according to the template from MDPI.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors use four well-known neural network architectures to solve aspect-based sentiment analysis. Related work section is strong. It was really interesting to see your contribution section that compared other related research. However, no new model or application is proposed. Also missing is a discussion section that explains the results.
Additional comments:
In the introduction, please highlight the main contributions of the paper.
Figures and tables should have more descriptive captions. There should be a key takeaway message conveyed in the caption.
Line #173, what exactly is "15.237 review data."
Table 3. A number of aspects have highly imbalanced data. How did it affect performance? Have you tried to balance it? Upon review of table 8, I see that it reflects the imbalance. However, with 31 samples, the CNN is able to achieve a f1 score of 0.8. This raises the possibility of an overfitting issue.
Figure 8. The accuracy curve around the fifth epoch appears to be confusing. However, the loss curve does not show the same decline. What led to the performance drop?
Figure 10. There is a sudden drop in accuracy at Epoch 80. Loss curves show the same pattern. How do you explain that?
Author Response
Dear Reviewer,
We appreciate your precious time and effort reviewing our manuscript and providing valuable comments. Your valuable and insightful advice has led to possible improvements in our manuscript. We have carefully considered and tried our best to revise all comments brought up by you as a reviewer. We hope the careful revision meets your high standards.
Below we provide the point-by-point responses. The original comments are provided in black color, whereas our answers are written in red.
We have also made corrections and refined the quality of the manuscript; these are also highlighted in yellow.
Sincerely,
Ahmad Abdul Chamid
Point 1: In the introduction, please highlight the main contributions of the paper.
Response 1: We thank you for your insightful suggestion. In the introduction, we have highlighted the main contributions to the manuscript in yellow. We've also highlighted the main contribution at point 2.1 Contribution and lines 156-164 in yellow.
Point 2: Figures and tables should have more descriptive captions. There should be a key takeaway message conveyed in the caption.
Response 2: We thank you for your insightful suggestion. We have added captions to the Figures and Tables as suggested; we highlighted them in yellow.
Point 3: Line #173, what exactly is "15.237 review data."
Response 3: We appreciate your comments to improve our manuscript. The scraping results obtained 15,237 data from product reviews on the Indonesian marketplace. In the manuscript, we have added an explanation of the data review; we highlighted them in yellow.
Point 4: Table 3. A number of aspects have highly imbalanced data. How did it affect performance? Have you tried to balance it? Upon review of table 8, I see that it reflects the imbalance. However, with 31 samples, the CNN is able to achieve a f1 score of 0.8. This raises the possibility of an overfitting issue.
Response 4: We also thank you for your valuable comment. In the process, aspects classification is not considered as features. All data is considered equal, so the difference in the amount of data does not affect. The process of grouping aspects is done without a training model but from the key phrases in the aspect sentences. So the difference in the amount of data does not affect the classification process.
Point 5: Figure 8. The accuracy curve around the fifth epoch appears to be confusing. However, the loss curve does not show the same decline. What led to the performance drop?
Response 5: We also thank you for your valuable comment. The use of adam optimization allows the model to escape local maximum accuracy. So from epoch 0 to before 5 the model is at a local minimum. There is no difference in the problem of loss because the loss calculation is carried out in a macro-average manner, while validation is carried out in binary.
Point 6: Figure 10. There is a sudden drop in accuracy at Epoch 80. Loss curves show the same pattern. How do you explain that?
Response 6: We also thank you for your valuable comment. In Epoch 80, due to the use of the Adam evaluation model, the stagnation in the previous epoch made the model try to move areas, fearing it would get stuck at the local maximum. At a loss, a model evaluation is also carried out to avoid a local minimum.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
All the problems have been addressed
Reviewer 2 Report
The authors have addressed all my comments.
