Explainable Turkish E-Commerce Review Classification Using a Multi-Transformer Fusion Framework and SHAP Analysis

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a framework for sentiment (useful or useless) classification of reviews using distributed representations derived from two types of BERT techniques. Given the rapid progress in machine learning and natural language processing, the proposed approach is considered valuable.
However, to make the paper complete, several points require further clarification, as outlined below.

<Major Comments>

1. A single review often consists of multiple sentences. While Tables 1 and 2 focus on short sentences, longer reviews may contain mixed content, including both useful and non-useful information. Please clarify how such long reviews were processed.

2. Concat Fusion achieved the best performance among the proposed methods. A possible explanation is that weighting features within the classifier provides greater flexibility than applying dimensionality reduction beforehand. Beyond reporting performance, the authors should discuss why this method performed best, in terms of model and data characteristics.

    

    

3. For the best-performing model, please analyze feature importance (e.g., using SHAP). In particular, when two distributed representations are concatenated for a single sentence, were both representations equally important?

<Minor Comments>

1. In the logistic regression analysis, were regularization techniques applied? Given the high dimensionality, multicollinearity may be a concern.

2.  

   If misclassified samples exhibit common characteristics, please discuss them.

Author Response

"Please see the attachment." 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1.Although the study used several models to make the classification of the text mining and sentimental analysis results, research hypothesis should be posited after the literature review part.

2.For the conclusion part, all results should be listed,  and authors should tell readers whether the results support the research hypothesis or not.

3.We suggest that authors should use cross validation in all models and make comparison.

   Besides, the confuse matrix presentation can be simplified by using the ROC curve for all models to make the comparison.

Author Response

"Please see the attachment."

Author Response File: Author Response.pdf