Methodology
How the Bangladesh Violence Tracker collects, classifies, and verifies incidents. We publish this so our data can be checked and cited.
Where the data comes from
Every few hours an automated crawler collects news from five major Bangladeshi outlets: Prothom Alo, Ajker Patrika, Jugantor, Samakal, and Dhaka Post. Each article is stored in full before any analysis, so the raw record is preserved.
When two or more outlets report the same incident, the reports are linked and the event is marked corroborated. Every event page lists all outlets that reported it, with links to the originals.
How an incident is classified
Each article is read by a language model with a fixed analyst prompt. It decides three things in order: is the event recent, is it a genuine act of violence, and did it happen in Bangladesh. Only if all three hold is a structured record created, with the location, date, parties, casualties, category, and a summary.
Articles that fail any test are kept in the archive with the reason recorded, so nothing is silently dropped.
The six categories
Political Violence
Party or student-wing clashes, election-related violence, rallies turning violent, attacks on party offices, violence during protests and hartals.
Criminal Violence
Gang violence, robbery, extortion, drug-trade violence, land or family disputes, and murders without a political or communal motive.
Mob Justice / Lynchings
Mob beatings or killings of suspected thieves, vigilante justice, and crowd violence against individuals.
Gender-Based Violence
Rape, sexual assault, domestic violence, acid attacks, dowry-related violence, and harassment that turns violent.
Terrorism / Extremist Attacks
Bomb attacks by militant groups, religious-extremist violence, and organised attacks on minorities by extremist groups.
Communal / Religious Violence
Clashes between religious communities, attacks on religious minorities, and desecration of places of worship.
What is excluded
The tracker records violence, not every death or disturbance. The following are deliberately left out:
- Road accidents, unless deliberate or part of an attack or blockade
- Natural deaths, suicides, drownings, and electrocution
- Deaths from lightning strikes and other natural causes
- Peaceful protests and processions with no physical altercation
- Court verdicts, press briefings, and threats with no action taken
- Simple arrests with no clash or shootout
- Fiction, literature, and opinion pieces
- Historical accounts of events from earlier years
How confidence is measured
Each record carries a composite confidence score, shown as a percentage. It combines several signals rather than a single guess: how many independent outlets corroborate the event, how complete the extracted fields are, how consistent the model's own reasoning was, and any community feedback on the record. A record checked by a human is marked Human-verified; one that readers have questioned is marked Community-flagged; the rest are AI-classified.
Community verification
Any reader can help verify a record. On each event you can say whether the classification, location, casualties, parties, and other details look right, and see how others voted. Signed-in users can suggest specific corrections. This feedback is aggregated, used to flag records for review, and released as part of the open dataset for research and for training better models.
Known limitations
- Coverage is limited to the five sources above; incidents reported nowhere are not captured.
- Automated classification makes mistakes, especially on ambiguous reports; this is why confidence, corroboration, and community verification are shown openly.
- Casualty figures reflect what sources reported at the time and may be revised later.
- Location is resolved to district level.
- Who-attacked-whom is harder to extract reliably than the fact of an incident, and is being actively improved.
Questions about the methodology or a specific record? Use the verification tools on any event page, or the correction button to flag an error.