- The paper introduces BOUTEF, a 57,095-post multilingual corpus covering Arabic dialects, Arabizi, French, English, comments, debunking content, genuine narratives, and 194 images from Algeria and Tunisia.
- The paper finds a strong association between themes and veracity (Cramér’s V = 0.49), with fabricated, misleading, and diverted content comprising 85.96% of fake items and political misinformation showing especially high engagement.
- The paper demonstrates that fake and genuine content have nearly identical normalized lexical entropy, while small, repetitive, rapidly spreading narratives produce the strongest echo-chamber effects, highlighting the limits of simple linguistic detection features.
Motivation and positioning
The paper introduces BOUTEF, a substantially expanded multilingual corpus for the study of fake news in Algeria and Tunisia, and presents a mixed-methods empirical analysis of misinformation in these two national contexts (2606.00193). The work is framed by the notion of attention war and cognitive war: fake news is treated not merely as false content but as a mechanism for capturing attention and shaping beliefs. The authors distinguish misinformation (unintentional inaccuracy) from disinformation (deliberate falsehood) and observe that politically motivated content in the corpus is most often disinformation, reflecting intentional manipulation that persists despite counter-evidence.
The corpus addresses a documented gap. Existing English resources such as FakeNewsNet, FEVER, and LIAR are large and well annotated, but Arabic resources remain limited, and prior Arabic datasets (AFND, DART, satirical news collections) do not jointly include user comments and debunking information. The authors also criticize a conceptual ambiguity in many existing corpora, which fail to separate genuinely fake content from texts that merely discuss or debunk it. BOUTEF is designed around three explicitly separated components: fake narratives, comments on fake narratives, and debunking (NoFake) content, complemented by genuine narratives and their comments for controlled comparison.
Corpus construction and composition
BOUTEF contains 57,095 posts and over 1.09 million words (104k unique word types), plus 194 images associated with fake content. The fake part is built by first identifying false claims through fact-checking outlets (Misbar, Falso, DZ Fake news, AFP), tracing those narratives on YouTube, Facebook, Twitter/X, and TikTok, and annotating each entry with language, country, gender, major/minor theme, and a fake category adapted from Wardle's taxonomy. Genuine narratives are retrieved using keywords derived from fake claims (e.g., verifying a claim about President Tebboune's death against reports of a later public appearance), with comments collected simultaneously.
The corpus spans MSA, Algerian and Tunisian dialects, Arabizi variants, French, English, Moroccan dialect, and code-switching. Acquisition is deliberately mixed: 50,297 items (88.1%) are SCRAPED genuine posts and comments, 6,280 (11.0%) are manually CHECKED items (fake narratives, fake comments, debunking posts), and 183 (0.3%) are FABRICATED items generated via few-shot LLM prompting and manually enriched with metadata. The fabricated subset is small but intended for stress-testing detectors on synthetic misinformation patterns. The corpus is publicly released on Hugging Face.
A notable structural imbalance deserves attention: within the full corpus, YouTube accounts for 89.07% of narratives, a direct consequence of the scraping pipeline, whereas the fake-only subset is dominated by Facebook (48.82%). The authors acknowledge that platform-aware normalization or stratified evaluation is necessary to avoid biased conclusions.
Thematic structure and statistical association with veracity
Thematic coding produced 67 themes. The distributions differ sharply between countries and between veracity classes. For Algeria, politics leads fake content (18.03% versus 12.58% for genuine), while genuine content emphasizes economy (24.50%), sports (13.25%), and culture (8.61%); fake content over-indexes on religion (5.80%), personality death (5.17%), and army (5.17%), with denigration appearing among major fake themes only. For Tunisia, war and economy tie as leading fake themes (14.46% each) followed by immigration (9.50%), while genuine content is dominated by economy (32.18%) and health and culture (11.49% each). In both countries, the residual "other themes" category is much larger for fake content (49.53% and 36.06%) than for genuine (34.45% and 17.25%), indicating that misinformation is thematically more dispersed while verified reporting clusters around public-information domains.
A chi-square test of independence between major theme and message label confirms a highly significant association (χ2=272.99, 34 degrees of freedom, p=4.12×10−39), with Cramér's V=0.49, a moderately strong effect. Residual analysis associates diplomacy, terrorism, economy, culture, health, and tourism with genuine messages, and denigration, humiliation, personality, racism, language, and personal illness with fake messages. The authors are careful to state this as association rather than causality, but the result directly supports theme-based feature engineering for detection models.
Within Wardle's categories, fabricated content dominates at 42.76%, followed by misleading content (24.46%) and diverted information (18.74%); together these three account for 85.96% of fake items, with exaggeration, manipulated content, satire, and denial forming a small tail.
Manipulation strategies and persistence
The qualitative analysis identifies recurring fabrication mechanisms: unconventional spellings and exaggerated punctuation to amplify outrage, out-of-context or altered images (over 22% of fake articles include visual media), sensational headlines, false authorities (e.g., a fictitious "European Institute of Verification"), and imitation of credible media formats. Emotional language combined with cherry-picking is illustrated with French and Arabic examples that blend factual elements with loaded framing. The paper's central qualitative claim is that fake content preserves a surface form mimicking credible reporting while activating fear, anger, and humiliation — a claim later corroborated quantitatively by the lexical entropy analysis discussed below.
A striking finding concerns longevity. The false claim that Neymar is of Algerian origin circulated for 3,533 days (nine years); the rumor of President Tebboune's death persisted 1,756 days; a story about a professor whose wife died in childbirth remained online 1,837 days despite rapid debunking. The authors conclude that debunking alone is insufficient, attributing persistence to confirmation bias, the backfire effect, echo chambers, and the speed asymmetry between viral spread and fact-checking documented by Vosoughi et al.
Engagement analysis shows a highly concentrated landscape: the top two fake-news accounts achieve comment-to-follower engagement rates of 96.00% and 95.60%, a second tier sits at roughly 26–45%, and the lowest observed rates (around 13%) are still non-negligible. Political fake news shows the strongest correlation between follower count and engagement, implying that large-audience creators amplify political misinformation most effectively. A documented case study — the fabricated report of Tebboune's death spread on TikTok by two influencers with 50,700 and 46,500 followers — illustrates how influencer distribution and algorithmic prioritization interact.
The paper operationalizes echo-chamber intensity at the narrative level with a composite Echo Chamber Index (ECI) combining semantic similarity (multilingual Sentence-BERT embeddings), author concentration (Herfindahl–Hirschman Index), and 24-hour temporal burstiness, weighted 0.5/0.3/0.2. Results show that small, highly repetitive, rapidly spreading narratives exhibit the strongest echo-chamber effects (e.g., ECI 0.87 for a death-of-Toufik narrative with a single author), while a high-volume narrative such as the Bouteflika death rumor (301 entries, 5 users) reaches only moderate ECI (0.55) due to near-zero burstiness. The implication is that message volume alone does not indicate echo-chamber structure; temporal synchronization and semantic redundancy matter more.
Sentiment and lexical diversity
Transformer-based sentiment classification (SieBERT for English, mBERT-based models for French, CAMeL Lab models for Arabic) yields 59.48% negative, 30.42% neutral, and 10.10% positive across fake content. The sizable neutral share suggests a pseudo-factual tone that may enhance credibility while conveying misleading claims — consistent with the media-imitation strategy identified qualitatively.
The lexical diversity analysis is the most methodologically interesting quantitative section. Using Yule's K, normalized Shannon entropy Hnorm​, and language entropy Elang​, the authors find:
| Type |
Posts |
Yule K |
Hnorm​ |
Elang​ |
| Fake |
891 |
18.07 |
0.90 |
1.53 |
| FakeComment |
5,142 |
14.40 |
0.86 |
2.56 |
| NoFake |
448 |
20.01 |
0.87 |
1.30 |
| Genuine |
309 |
26.07 |
0.90 |
2.02 |
| GenuineComment |
50,321 |
15.91 |
0.76 |
1.75 |
Two results stand out. First, Fake and Genuine messages share an identical normalized entropy (0.90) and nearly identical per-message entropy medians (~3.9 vs. 4.1 bits), with overlapping notch confidence intervals; only NoFake is statistically separable at the 5% level. This is direct quantitative evidence that fake news mimics the lexical distribution of legitimate reporting, and it implies that per-message lexical entropy is insufficient as a standalone detection feature. Second, language entropy separates producers from audiences: fake posts are relatively linguistically homogeneous (Elang​=1.53) while the comments they attract are the most multilingual content in the corpus (Elang​=2.56), and debunking content is the most monolingual (1.30), reflecting fact-checkers' deliberate choice of MSA and French. The authors interpret this as strategic language selection by misinformation producers versus spontaneous, linguistically diverse audience reaction — a sociolinguistically meaningful asymmetry, though it is an interpretive claim rather than a demonstrated causal one.
Geographic and demographic patterns
Source attribution differs markedly by target country. For fake news targeting Algeria, unknown origins account for 56%, with Morocco (17.5%) and Algeria itself (16.3%) the largest identified sources, plus small shares from France, Egypt, and Israel — a transnational profile. For Tunisia, unknown origins account for 36.7% but Tunisia itself dominates identified origins at 55.3%, indicating a predominantly domestic misinformation ecosystem. The authors correctly flag the large unknown share as a limitation on source attribution.
Demographic analysis is weaker. Gender is undetermined for a large share of creators (45.9% overall), and within the identifiable subset, men are more associated with politics (19.13%) and sports (8.69%) while women appear more in immigration-related fakes (11.76%). The paper acknowledges that age cannot be verified and that representation is uneven, so these findings should be treated as indicative only.
Limitations and open questions
The paper concedes several limitations that qualify its results. The large proportion of unattributed sources (UNK) limits geographic conclusions; the YouTube-heavy composition of the full corpus risks platform bias; the fabricated subset is very small (183 items); and the corpus lacks user age information, weakening demographic inference. The ECI weights (p=4.12×10−390, p=4.12×10−391, p=4.12×10−392) are chosen rather than learned, and the subject field is used as a proxy for narrative clustering, both of which are assumptions a reader should bear in mind. Open questions left by the paper include: whether diversity-based features (p=4.12×10−393, p=4.12×10−394) improve detection models in combination with multimodal representations; how echo-chamber dynamics generalize to other Maghreb countries; and whether counter-narratives are effective against the long-persistence narratives the corpus documents.
Conclusion
BOUTEF provides a publicly available, annotated, multilingual resource that uniquely combines fake narratives, user comments, debunking content, and genuine comparators for Algerian and Tunisian social media discourse. Its principal empirical contributions are a statistically strong theme–veracity association (Cramér's p=4.12×10−395), evidence of extreme narrative longevity despite debunking, a demonstration that lexical entropy cannot separate fake from genuine news, and a composite echo-chamber index showing that small, synchronized, repetitive narratives exhibit the strongest echo-chamber dynamics. The comparative analysis establishes that misinformation targeting Algeria has a more transnational origin structure than the domestically rooted Tunisian ecosystem, though attribution uncertainty tempers this contrast. The corpus and its analyses offer a concrete empirical basis for detection research in under-resourced Arabic dialects and for studying information disorders in multilingual environments.