Research · Accepted at ASONAM 2026

A Social and Legal Discourse Analysis of Women's Safety Narratives in India

Yogesh Kuchimanchi · Charishma Reddy Yerra
Rochester Institute of Technology

We studied 351,501 retained Reddit and YouTube comments across 16 cases in India, with event years from 2012 to 2024. These charts show how the discussions differed by case and over time.

Government criticism by case

Share of retained case comments classified as critical of government, sorted from highest to lowest.

Showing 16 of 16 cases
Political implicationNon-political
Sandeshkhali 20243,720 comments · Political
48.55%
Hathras 202014,297 comments · Political
40.71%
Unnao 20179,113 comments · Political
37.08%
RG Kar 20249,981 comments · Political
36.27%
Manipur 202315,250 comments · Political
30.52%
Hyderabad Vet 20195,194 comments · Non-political
23.97%
Kathua 201814,082 comments · Political
23.51%
Bhavana Assault 20171,634 comments · Non-political
23.26%
Nirbhaya 201235,517 comments · Non-political
19.34%
Badaun 20143,424 comments · Non-political
17.61%
Badlapur 20245,937 comments · Non-political
16.91%
Shraddha Walkar 20221,064 comments · Non-political
16.35%
Jisha 20161,261 comments · Non-political
16.10%
Shakti Mills 20132,898 comments · Non-political
13.11%
Uber Delhi 20142,637 comments · Non-political
12.36%
Pollachi 20194,849 comments · Non-political
11.71%
The submitted paper reports a median of 36.67% for political cases and 16.65% for the non-political comparison (Mann–Whitney U = 59, p = 0.0005, r = 0.786; a large case-level effect). Filtering changes the displayed cases, not this reported comparison.
View case data
Values shown in the filtered chart
CaseGroupCommentsGovernment criticism (%)
Sandeshkhali 2024Political3,72048.55
Hathras 2020Political14,29740.71
Unnao 2017Political9,11337.08
RG Kar 2024Political9,98136.27
Manipur 2023Political15,25030.52
Hyderabad Vet 2019Non-political5,19423.97
Kathua 2018Political14,08223.51
Bhavana Assault 2017Non-political1,63423.26
Nirbhaya 2012Non-political35,51719.34
Badaun 2014Non-political3,42417.61
Badlapur 2024Non-political5,93716.91
Shraddha Walkar 2022Non-political1,06416.35
Jisha 2016Non-political1,26116.10
Shakti Mills 2013Non-political2,89813.11
Uber Delhi 2014Non-political2,63712.36
Pollachi 2019Non-political4,84911.71

Stance across time windows

The baseline is a separate comparison corpus. The other rows group comments by time since the case.

Critical of governmentCritical of societyNeutral reportingSupportive of victim
General baselinecomparison corpus
Acute0–90 days
Sustained91–365 days
Retrospective366+ days
The retrospective row shows a lower government-criticism share than the acute row. The paired acute-to-retrospective test across 15 eligible cases did not find a statistically significant change (Wilcoxon W = 43, p = 0.359). Each eligible case had at least 20 rows in both windows.
View time-window data
Stance shares (%); rounding may affect row totals
WindowCritical of governmentCritical of societyNeutral reportingSupportive of victim
General baselinecomparison corpus13.85957.04228.5990.500
Acute0–90 days33.86226.09336.5363.510
Sustained91–365 days29.29329.59737.4823.628
Retrospective366+ days23.93830.16541.1444.753

Recorded institutional responses

This taxonomy describes the responses recorded in the paper. It does not show that online discussion caused legal or policy action.

Response categories across the 16 cases
ResponseCasesExamples / coverage
Legislative2Nirbhaya 2012 · Kathua 2018
Regulatory / state action3Uber Delhi 2014 · Hyderabad Vet 2019 · RG Kar 2024
Executive / judicial only11Remaining cases in the paper taxonomy

Later model audit

These results use a fixed 3,000-row validation matrix from a post-submission local audit. They are separate from the submitted paper’s pilot results.

BaseAdapted

Accuracy

Base68.00
Adapted74.47

Macro-F1

Base67.54
Adapted74.50

MCC

Base59.81
Adapted66.34
The adapted Qwen model scored higher on all three measures. Its high-confidence error rate also rose, from 10.51% to 11.74%. Better classification did not mean better calibration in this audit.
View model-audit data
Reported scores from the later validation matrix
MeasureBaseAdapted
Accuracy68.0074.47
Macro-F167.5474.50
MCC59.8166.34
High-confidence error rate (%)10.5111.74

Methods and limitations

Public Reddit and YouTube comments were collected, cleaned and classified into four stance labels, then compared by case and time window. The accepted manuscript reports 351,501 retained comments, 16 cases with event years from 2012–2024, and 92.8% annotator agreement.

The corpus summary and case/time comparisons above retain the submitted-paper figures. The later model audit uses local reproducibility artifacts; it does not replace the submitted pilot.

  • Platform participation is not representative of India as a whole.
  • Observed discourse–response patterns are associations, not causal estimates.
  • Only aggregate statistics are shown. Raw usernames, comment text and individual-level records are not published here.