Aspect-Based Sentiment Analysis with Dependency Relation Weighted Graph Attention

Round 1

Reviewer 1 Report

The authors then propose a new approach for sentiment analysis using graph neural networks (GNNs) combined with syntactic analysis, to overcome limitations of traditional neural networks in processing data in Euclidean space. The paper presents a model for aspect-based sentiment analysis using sentiment pre-trained dependency-weighted graph attention networks, with the aim of extracting graph structural features, weighting the attention weights according to dependency weights, and improving the prediction accuracy and F1 values of the model.

My concerns are as follows:

1-The introduction does not seem to provide a comprehensive overview of the background literature or all relevant references. It is also not clear what prior work the authors are building upon or how their work differs from previous approaches. It would be helpful to include more context and references to provide a clearer understanding of the motivation and contribution of the paper. 

2- What are the three publicly available aspect-based sentiment analysis datasets used in this paper?

3- What is the difference in word embedding dimensions between Glove and BERT pre-trained models used in this study?

4- What are the evaluation metrics used in this study to evaluate the model results?

4- What are the 11 baseline models used in this study for comparison with the proposed model?

Limitations:

1- The paper only uses three datasets for experimentation, which may not be enough to generalize the results for other datasets or real-world scenarios.

2- The paper does not discuss any limitations or challenges in the proposed model, which may limit its application in practical scenarios.

3- The paper does not compare the results with other state-of-the-art models, which makes it difficult to assess the significance of the results in the context of other existing models.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

I would first like to thank the authors for their submission and for the privilege to review their work. Overall, I believe that the study may present a novel contribution to the extant field of research. However, there are several limitations that need better clarification in the presentation of the manuscript and its contents.

1.) Most notably, the contribution of the research is not readily apparent to the reader, and thus, I strongly urge the authors to highlight their study's contribution to the extant body of literature on aspect-based sentiment analysis.
1a.) Specifically, in the abstract, the authors try to point to the relevance of their research in lines 22 and 23, wherein "WGAT model had better performance in terms of accuracy and F1 values." However, the authors should better specify what the proposed model performed better compared to. In this instance, it is notable that the WGAT model performed better than about a dozen other existing top of the line models. This would immediately demonstrate the notability and contribution of the proposed model.
1b.) At the end of the Introduction section (lines 63~87), the authors explain extensively about the methodology of the proposed model, and allude to the fact that these are contributions, however, it is unclear where the actual notable contributions to theory is within these paragraphs. It is recommended to more explicitly point out the contributions of this research study, rather than merely describe the steps of the method. For example, it seems that the proposed method is an improvement on extant methods in that it is able to more accurately consider the importances of different types of dependencies in assigning weights. The unique contributions should be highlighted towards the end of the introduction in order to clarify why this method is novel and worth attention.
1c.) Finally, the contributions of this study are not mentioned in the conclusion, and should be reiterated in the conclusion. 

2.) There are some aspects of the methodology that should be clarified more precisely in the manuscript.
2a.) There is little mention of preprocessing of the text prior to modeling. What preprocessing steps were taken to clean the text data, if any?
2b.) Some areas of the manuscript require further detail. For example, on line 180, the authors define A as the "n aspect words in the sentence." However, there does not seem to be a clear definition of which words constitute as aspect words and which do not. Furthermore, are stopwords taken out (see comment 2a), and if not, would these affect the number of aspect words in a sentence? In short, please clarify any definitions within the methodology more precisely.
2c.) Little information is given about the cross validation mentioned on lines 329 and 330. While a table is provided, it is unclear what decisions were made in order to create the training and testing sets. Please specify more clearly whether the train/test splits were randomized, stratified by sentiment, etc. While the reader may assume that the splits were randomized, it should be more clear in the text.
2d.) Lines 472~473, the authors state that "the control variables method" was used, which is unclear as to which "control variables method" the authors are referring to. If the authors are simply referring to using control variables, then it would be recommended not to use "the" which implies that it may be a specific method for finding a solution, rather than a standard technique as part of a larger method. However, if the authors are referring to a specific control variables method used in similar studies, then a citation might be in order.

3.) Figure 5 refers to two dependency weightings, one labeled a and one labeled b. Within the text, this becomes confusing the a and b are not denoted with special characters or italics. It is recommended to better clarify the a and b from the normal text via italics or parentheses.
3a.) It would also be more clear if the subheading of figure 5 were to denote how a and b differ, so as the reader does not have to search for the answer in the text. For example (although it does not need to follow this exact wording), 'Performance of WGAT model on each dataset under important dependency weights (a) and other dependency weights (b)'. 
3b.) The y axes are not aligned so it is difficult to directly compare 5a and 5b. It is recommended to set the y-axis limits to the same values so that the two figures can be directly compared.

4.) It is recommended to expand on the Conclusion section in more detail. As mentioned in a prior comment, the authors are strongly urged to focus more on the contribution of their paper to the theoretical literature and practical implications. However, the authors should also be more specific about the limitations of the model. For example, the authors mention that "...it faces some special sentence types with prediction errors", but the authors should briefly point out which sentences cause these kinds of errors and why. The goal of pointing out the limitations is to focus researchers on problems that still need solving, thus, the authors should be more clear what limitations these are.

5.) There are several minor spelling mistakes or other typos that need correction, which include but are not limited to the following:
5a.) Line 306, the authors denote sigma as the activation function, "where  denotes the number of attention heads". However, there is a character missing here. Please check this.
5b.) Line 474: 'dataset' is misspelled 
5c.) Line 523: the word 'Please' is unnecessary, please delete.
5d.) Line 525: 'conflict' is misspelled

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have answered my concerns. I recommend the acceptance of the manuscript on its current form.