Wikipedia’s accounts of war can portray a language community’s own side as more powerful or capable than its enemy — but the direction of that difference changes from one language edition to another, according to a new analysis of 16,058 articles.
English, Japanese, Hebrew and Russian articles showed significantly stronger portrayals of self-aligned combatants on power or agency. Arabic, Ukrainian and Vietnamese articles showed significantly weaker self-aligned portrayals on those measures. Russian differed across all three dimensions.
The findings come from a dataset covering 2,542 battles in 158 wars fought since 1900, across 20 Wikipedia language editions. The researchers measured how combatants were represented through three dimensions: power, agency — whether an actor is represented as able to act — and sentiment.
The gap was not simply about allies
The allied-versus-enemy test did not find a comparable pattern. After correction for multiple comparisons, no language showed a significant difference on any connotation dimension. Ukrainian and Hebrew were excluded because their samples were too small.
To make the comparisons, the researchers used software that identifies agent–verb–theme triples — who did what to whom — and assigns scores based on the connotations of those relations. The self-enemy comparisons were then tested with a paired statistical method.
The final corpus was narrowed using a rule requiring at least 10 edits. That removed 1,288 observations, or 7.4% of the original sample, leaving 16,058 articles and reducing the number of battles from 2,581 to 2,542.
Words of resistance, failure and loss
The language of the articles also shifted with the direction of the framing. Topic keywords linked with higher self-aligned power or agency scores included descriptions of defense and resistance. Lower-score descriptions, in several languages, were associated with words about inability, failure or casualties.
The study found that battle outcomes were associated with these gaps in the text. Victories were linked to wider differences in power, agency and sentiment between self and enemy portrayals, while losses were linked to smaller power and agency differences. These are associations in the article data, not evidence that winning or losing caused the wording.
Editing patterns were also related to how enemies appeared. Greater concentration of edits and higher revert rates were associated with weaker enemy-agency portrayals. Articles with more unique editors were associated with higher power scores for both sides and more positive enemy sentiment.
Different routes to a similar imbalance
The statistical models suggested that similar self-enemy gaps could be produced in different ways. Japanese articles combined stronger portrayals of the self-aligned side with weaker portrayals of the enemy, while the gaps in Hebrew and Russian articles mainly reflected stronger portrayals of the self-aligned side.
Across the 20 language editions, higher win rates were associated with larger power and agency gaps, while higher loss rates were associated with smaller gaps. A cultural measure called Long-Term Orientation was also positively associated with both gaps. The study found no significant relationship for sentiment gaps or for the remaining cultural and editorial measures it tested.
To test whether the pattern depended on the main regression setup, the researchers ran a mixed-effects model that accounts for language-level clustering. It produced substantively similar results, but the regression findings remain comparative associations rather than causal effects.
Non-involved editions often moved together
The picture changed when the language community was not directly involved in a war. For those conflicts, the editions showed high narrative similarity, with scores ranging from 0.7 to 0.9. A clustering analysis found no clear separate language blocs, although some editions occupied more central positions in the network and others were more peripheral.
High similarity may partly reflect sparse editing histories or other data limitations. The study applied a 10-edit threshold, but the convergence results were not separated according to how much each article had been edited.
A broad computational measure with clear limits
The researchers translated non-English articles into English and back-translated English articles through Spanish. A replication using another translation system produced comparable results, although translation can blur culturally specific meanings or introduce artifacts.
Human validation of 50 subject-verb-object triples in English, Arabic, Japanese and Chinese found no significant cross-language difference in agreement for power or agency, but lower consistency for sentiment. Sentiment findings therefore need particular caution.
The paper describes its language-level correlations as exploratory, and the analysis used only 20 language editions, limiting statistical power for smaller groups. The study is also comparative and descriptive, so its results do not establish that cultural values, editorial behavior or battle outcomes caused the framing differences.
The paper is an arXiv preprint, version 1, dated 26 August 2026. The translation work received computing credits from the Google Cloud Research Credits Program.
Paper data and sources
Original title: Framing War Across Languages: Power, Agency, and Sentiment in Wikipedia's Multilingual War Narratives
Authors: Jiarui Xia, Diego Gomez-Zara
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text