📄 Pursing power in Arabic on-line discussion forums

Summary

The paper presents an Arabic corpus for spotting forum participants who try to raise their informal standing in a group. The focus is on expert power or respect from peers, not formal rank. Instead of guessing how someone seeks influence, the authors rely on linguistic signals of conflict and status.

The data are MSA discussions from Arabic WikiTalk pages and reader comments on Al Arabiya. Both were filtered so that the threads have similar levels of interaction. Five native-speaker annotators worked at two levels:

• Person level: each participant got a graded judgement of whether they pursue power.
• Sentence level: sentences were tagged with attributes tied to ten "social acts". These acts draw on Keltner et al.'s interactive model of power and on theories of group conflict.

About a third of participants were judged to be seeking power. Agreement on the person-level labels was about 82%. A transition graph of social acts showed patterns that fit the psychological theories and also differ from English data. The resource supports modelling upward social movement and cross-cultural comparison.

Key points

• Ten psychologically motivated social acts, realised as 36 sentence-level attributes. Examples include establishing and challenging credibility, agreement and disagreement, task and relationship conflict, and leadership avoidance.
• Pursuit-of-power labels used a scale from −3 to +3, with 0 meaning uncertain. Counts: 3196 annotations, 2424 double-annotated, 82.17% sign agreement, 59 data units. Attribute labels: 770 annotations, 141 double-annotated, 95% agreement.
• Thread selection:
  • WikiTalk pages had 3–8 speakers and roughly 14 turns.
  • News comment chains were kept when 2–5 speakers each posted more than three times, totalling 12–19 turns.
• Inter-annotator results: the chance that a second annotator also gave a positive label rose with the first annotator's confidence, from about 0.18 to 0.65 to 0.77. Binomial tests found these differences significant at α = .05.
• An English comparison graph was built from 70 WikiTalk discussions. In Arabic, direct credibility challenges followed by group affordances signalled power-seeking more strongly than in English.

Interesting ideas

• I1. Graded labels with confidence: a signed confidence scale, with agreement scored only on the sign, lets you correct for each annotator's personal threshold and check whether stated confidence actually predicts agreement.
• I2. Mid-level social acts: placing a layer of social acts between surface sentences and a high-level judgement (who seeks power) keeps the final label traceable to concrete utterances.
• I3. Significance-pruned transition graph: model the dialogue as a chain of acts and keep only transitions that pass a chi-square test, showing each node's strongest positive and negative edge. This gives a readable, explainable predictor.
• I4. Interaction-matched sampling: filter two very different sources by number of active speakers and total turns so they become comparable before annotation.
• I5. Undefined attributes plus drift checks: attributes were deliberately left without formal definitions to allow language-specific variation. Drift was controlled by periodically translating annotations and discussing them.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Witness parsers or LLM teacher tiers could give signed confidences, and silver data could be kept only where high confidence is shown to predict agreement.
• I2 (frames): Crate prag could add social acts such as challenging credibility or explicit disagreement next to speech acts, so judgements about characters' intentions cite specific clauses.
• I3 (world): Sequences of events or acts in the world of a text could be summarised as pruned transition graphs that explain why a character is perceived a certain way.
• I4 (learn): When building evaluation sets from mixed genres, match on structural statistics so that score differences reflect the model rather than the source.
• I5 (learn): LLM-teacher labels could be sampled and reviewed regularly against reference examples to detect drift from the intended category.