An Enhanced Neural Word Embedding Model for Transfer Learning

Round 1

Reviewer 1 Report

The authors present a methodology to create word embedding (based on fastText) model for Bangla and use it for text classification purposes afterwards.

While the general methodology is generally sound (although it can be seen in many similar articles nowadays), the article must be greatly improved before it can be accepted.

1. The article contains a very large number of grammatical and spelling mistakes. There are also problems with respect to wording. This is not acceptable, especially for an article dealing with NLP. Please use (at least) services like Grammarly to fix these problems. Proofread your article before submitting! Examples: "the proposed work accomplishes text classification based on three popular textual Bangla datasets, developed models on top of that various machine learning classical approaches as well as deep neural network" ; "this work has been utilized a collected"; "this paper has been proposed"; "work has been used different"; "has been coded"; "Since these models are proposed a BanglaLM dataset,"; "with the be the maximization"; "data is decorated to a structure"....
2. Do not use ponouns like "we" and "our". Instead, use "the authors"....
3. What do you mean by "20 million observations"?
4. There is no subsection 3.2. Why is there a subsection 3.1 then?
5. How does the trained BanglaFastText model relate to the downstream text classification experiments? Will it be applied before the real ML model is used (on the three datasets), e.g. before the embedding layer of the CNN/LSTM, in order to get the word embeddings for the input texts? Why do you use different datasets for the creation of the BanglaFastText model and the text classification tasks?
6. The setup of the CNN and LSTM is badly described: You only explain how you set the respective architecture up, but not WHY you use for instance two LSTM layers. A clear explanation for your design choices is missing. Maybe a reference to a similar architecture from literature could be helpful...
7. Section 5.5 (too late) should become section 5.2 . A bullet is missing before the third dataset. What is the Mendeley dataset? I suppose, it is the second one, right? Please consistently use the same identifiers.
8. The main text classification tasks are not described for the different datasets. Please explain clearly what you have evaluated for which purpose using the input texts.
9. You mention "Seaborn and matplotlib have been used for data visualization". Where is a data visualization here?
10. Tables like table 4 should not be the last element in a section.

As you can see, your article contains many problems as of now and therefore cannot be accepted as is.

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

First of all, I would like to congratulate the authors on a very interesting and valuable contribution to this topic. Nevertheless, there are a few points that I would like to highlight:

The manuscript needs extensive revision for language and grammar since some parts were difficult to understand.

In the section “2.1. Word embedding of various languages” it will be interesting to note which languages each author studied (was it only English?) as well as to go a little deeper into the results obtained (stating, for example, the numbers of words or the corpora used, the accuracy percentage…).

In the section “2.2. Bengali Word Embedding” you claimed that “Ritu et al. [20] analyzed the most commonly used Bengali word embedding models”, which models did they use?

In the section “5.6. Comparison with Previous Techniques” I think it would be beneficial to add the results obtained in other languages using fast text word embedding methods (such as the research mentioned in the section 2.1.) and compare them with research in Bengali and with your results.

The conclusions section is too short. In my opinion, more critical information can be extracted from the authors' work. Limitations should also be included.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The article now indeed has improved.

Please use anyhow a spell checker again. Likewise, not all occurrences of "we" and "our" have been removed.

The article's title mentions an "enhancement", but this enhancement is not reflected in the methodology or obtained results AND it is not referred to in the article again. Please think about this again.

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf