Reviewer #1:

This is an interesting article and would be of interest to readers. However the
manuscript would be improved with reorganizing with the background section focusing
on the problem (digital harassment/threats) globally, why focus on LMIC, specifically
India, significant use of twitter vs. other LMIC, English speaking, social and gender
norms that reinforce inequality, etc) consequences of digital harassment/threats on
women/girls, specific details on COVID-19 in context and why important to examine the
association and then move to the theory/framework and categories that guides the
research and the study purpose and question.

The methods section is dense, perhaps divide into phases/sections related to 1.
geotagging locations of 30 million tweets (also how did you get access to the tweets for
analysis), 2) Qualitative analysis of tweets for misogyny and non-misogyny (curious how
students trained), 3. Test/Train models to determine best fit and 4) Analysis of misogyny
and COVID in the digital space using model.

The most lacking section of the paper is, what are the implications of these findings?
What should/could be done to prevent/reduce digital abuse/harassment during a
national/global crisis like a pandemic? What the policy and future research implications?

Reviewer #2:

In general the authors show competence, for this is an interesting and well-developed
study. However, I was left with some question and recommendations that I'd like to see
answered in the paper itself. That said, I believe this study should be published, once
these concerns are addressed.

- When describing the hypothesis and conclusions, these are worded as universal but
the study is restricted to tweets from India.

- Why the authors didn't explore modern NLP classifiers/techniques? We see TF-IDF
unigrams used with SVMs, bayes, logistic regression and a multi-layered perceptron.
Nothing wrong with those, but there are plenty of newer architectures with better results.

- My mayor concern here is that I missed a connection, and to me this work reads as
two papers. First, a taxonomy for misogyny is proposed and used for the classifiers,
then this classifier is used on tweets to measure said misogyny. But I missed the
connecting idea between the two. If the goal is to defend the proposed taxonomy, I'd
expect a comparison against previous taxonomies. If the goal is the time-series
analysis, I don't know why the previous taxonomies weren't enough to make the
measurements. Thus I read this work as having those two separate goals, and each has
its well-earned merits, but it lacks a paragraph somewhere connecting these two ideas.

Small notes on form:
1. Line [147] the figure is missing

2. Line [311] this conclusion is not constraint to India, while the study was

3. Line [319] can we justify this with most being less than 1% of the ground truth tweets