Manuscript Number: YCSLA-D-21-00263  

Offensive Language Identification in Low-resourced Code-mixed Dravidian languages using Pseudo-labeling

Dear Dr. Chakravarthi,

Thank you for submitting your manuscript to Computer Speech & Language.

I have completed my evaluation of your manuscript. The reviewers recommend reconsideration of your manuscript following major revision. I invite you to resubmit your manuscript after addressing the comments below. Please resubmit your revised manuscript by Mar 03, 2022.

When revising your manuscript, please consider all issues mentioned in the reviewers' comments carefully: please outline every change made in response to their comments and provide suitable rebuttals for any comments not addressed. Please note that your revised submission may need to be re-reviewed. 

To submit your revised manuscript, please log in as an author at https://www.editorialmanager.com/ycsla/, and navigate to the "Submissions Needing Revision" folder.  

Computer Speech & Language values your contribution and I look forward to receiving your revised manuscript.

Kind regards,    
Srinivas Bangalore  
Editor  

Computer Speech & Language

Editor and Reviewer comments:



Reviewer #1: The language identification of code-mixed languages is a practical research, and this paper proposes an approach to  solve this problem by generating pseudo-label. This paper is well writtened, and examples are clarify. However, there is only a few paragraphs showing how they get pseudo-label nor how they use the pseudo-labels to get performance improvements. And there are too many introductions of different BERT models. This paper looks not much innovative, and I suggest the authors highlight more on new approaches. The results show performance improvements but this seems to the attribution of data increasing.



Reviewer #2: 1. Authors mentioned that "Several researchers have worked on developing systems to detect the offensive language in English, Turkish
(Zampieri et al., 2020)". Actually, this shared task was for 5 languages not only two.

2. Some parts in "Dravidian Languages" section need examples and explanation, ex: where *ezuttu* is defined

3. A small typo: insert a space before "(" as in Tigalari script(Sekhar, 1951;..

4. Missing "(" in: Offensive-Targeted-Insult-Individual OTI)

5. I think it will be useful to add percentages to tables 2 and 3

6. Please add more explanation to the last 2 rows in figures 1, 2 and 3

7. You can make some numbers bold in table 5 to show best results.

8. MOST IMPORTANTLY, I couldn't understand the main idea of the paper of using pseudo-labeling.
Starting from a small number of annotated examples, how can we increase training data?
If we want to apply this to other languages, ex: Urdu, what are the exact steps that we can follow?
I read the paper carefully, but found this crucial part hard to understand.