📄 Linguistic Uncertainty and Engagement in Arabic-Language X (formerly Twitter) Discourse

Summary

The paper asks whether tweets that sound unsure draw different audience reactions than confident ones. It also asks whether any difference lies in the amount of engagement or in its type. The author used Apify to collect 16,695 deduplicated Arabic-language tweets mentioning Lebanon or Beirut over 35 days (December 2025 to January 2026). Uncertainty was detected with an interpretable Arabic lexicon of 60 markers in six categories. Context rules suppress false hits. For example, the particle "من" counts as "who" only when it is interrogative, not when it is the preposition "from". Roughly 30% of tweets were flagged as uncertain. These tweets had 51.5% higher mean engagement. A regression on log(1 + engagement), controlling for length, links and verification, gave about 25% more expected engagement. The effect was largest for replies, then retweets, then likes. The author names this the Uncertainty–Reply Asymmetry. The study matters for two reasons. It offers a cheap, transparent way to measure epistemic stance in an under-studied language. It also shows that engagement should be analysed by type rather than as one total.

Key points

• Data: 17,343 tweets were scraped with the query for Lebanon/Beirut and lang:ar. Removing 648 duplicates left 16,695 tweets, of which 6,872 (41.2%) are replies.

• Lexicon: 60 markers in six categories:

  • modal verbs (6)
  • hedges (12)
  • questions (7)
  • contextual who-questions (18)
  • information-uncertainty phrases (9)
  • rumor markers (8)

  The lexicon and rules were frozen before analysis and released as supplementary material.

• Labels: 4,997 tweets (29.9%) were labelled uncertain and 11,698 certain. A tweet counts as uncertain if it contains any marker.

• Validation: one native Lebanese Arabic annotator labelled a stratified sample of 200 tweets.

  • Accuracy was 73.5% and recall 1.00.
  • Precision was only 0.47, giving F1 0.639 and Cohen's κ 0.470.
  • The classifier therefore over-predicts uncertainty.

• Main result: β = 0.221 (SE 0.044, p < 0.001) on log engagement. The result holds under alternative specifications and author-clustered standard errors. The design is observational, and follower count, posting time and algorithmic amplification are not controlled.

Interesting ideas

• I1. Ambiguous lexicon entries fire only in a licensing context, such as a function word counting as a question word only when it is interrogative. This turns a crude word list into a usable detector at almost no cost. — For Mova (frames): hedge and question detection in prag could gate cue words on UD structure instead of strings, e.g. "may" only as aux, "who" only as a wh-word in an interrogative clause.
• I2. Each marker category is tied to a specific linguistic function: epistemic possibility, speaker doubt, information seeking, agent/source identification, or unverified report. Detection output thus says why something is uncertain, not just that it is. — For Mova (world): facts entering the world of a text could carry a certainty tag by source type (hedged, asked, rumored), so the answer step does not treat rumors as established states.
• I3. The lexicon and rules were fixed before any outcome analysis, which guards against quietly tuning the detector toward the hoped-for result. — For Mova (learn): self-induced repair rules and cue lists could be frozen before an evaluation run and checked against negative controls, so improvements are not artefacts of tuning on the test.
• I4. The validation sample was balanced across the two predicted classes, which exposed the pattern of perfect recall with weak precision. That pattern suggests using the rule set as a high-recall candidate generator followed by a stricter filter. — For Mova (learn): rule-based components could be evaluated on balanced predicted-positive and predicted-negative samples, and high-recall rules paired with a tree-based second-stage filter.
• I5. The outcome is split by type (replies vs. retweets vs. likes) rather than summed, which reveals effects that a single aggregate hides. — For Mova (answer): reader and student scores could be reported per question type (who/why/how-many, inference vs. lookup), so a gain in one type cannot mask a loss in another.