How MeydaNews works

MeydaNews combines news collection, textual and semantic similarity, and automated analysis to compare reporting. This describes the current approach and its limitations.

1. Collection and country editions

Reports are collected from configured, enabled sources. A country edition identifies the source collection, not necessarily the subject of every article. Fetch failures, access restrictions and publication delays can leave sources missing.

2. Grouping reports

The pipeline uses semantic embeddings, headline similarity and topic/event rules to group potentially related reports. Live clusters can change during processing. Similar events may be merged incorrectly, or one event may be split across clusters.

3. Cluster confidence

Cluster confidence is a heuristic derived from cluster characteristics, including article counts. It is not a calibrated probability, a publisher credibility rating or a verification of the reported event.

4. Claims, evidence and translation

The dashboard may automatically extract claims and textual evidence relationships. Support and contradiction labels describe the system’s interpretation of reporting, not independent factual adjudication. Translations may lose nuance; consult the linked original text.

5. Public, citable examples

Comparison pages are selected from groups with at least two named sources. Headlines, links and reporting timestamps are preserved in a dated snapshot. Selection does not certify accuracy or source independence. These pages are not live news updates and do not publish extracted claims as verified facts.

6. Dates and corrections

The snapshot date is separate from publishers’ publication dates. Displayed times use UTC. Snapshot URLs remain stable; corrections receive a dated change note. Missing publication times are explicitly marked unknown.