📄 Thematic Dispersion in Arabic Applied Linguistics: A Bibliometric Analysis using Brookes' Measure

Summary

The paper measures how concentrated or spread out recent research on Arabic Applied Linguistics is across its subfields. From Scopus, the authors gathered publications from 2019–2025 using a Boolean query. They screened the results by hand and placed each of 1,564 papers into exactly one of eight sub-disciplines, checking coding consistency with a second rater. They then computed Brookes' Measure of Categorical Dispersion, Δ = (M − 1)/(N − 1). N is the number of categories and M is the frequency-weighted mean rank, where the largest category gets rank 1 and ties share an averaged rank. The result is Δ = 0.194, which the authors read as strong dispersion. Computational Linguistics leads, but Sociolinguistics, Language Teaching and other subfields remain substantial. The authors attribute this pluralism to the complexity of Arabic: diglossia, dialect variety, a distinctive script and rich morphology. They also position the work as a reusable template for applying Brookes' formula. Readers should note that the paper's description of the extremes (which M gives Δ = 0 or 1) and of the direction in which the long tail moves M does not fit cleanly with its stated ranking rule. The concluding restatement of the formula also contains a typo.

Key points

• The dataset has 1,564 Scopus-indexed publications from 2019–2025, manually cleaned of false positives, non-linguistics papers and duplicates.
• The eight mutually exclusive categories are:
  • Computational Linguistics/NLP
  • Sociolinguistics
  • Language Teaching
  • Discourse Analysis
  • SLA
  • Applied Linguistics (General)
  • Corpus Linguistics
  • Language Assessment
• A second rater independently recoded a stratified 15% sample (about 235 papers), giving Cohen's κ = 0.87.
• The measured value is Δ = 0.194. The four smallest categories (SLA, Corpus, General, Assessment) form a long tail of roughly 13% of output.
• Δ is proposed as a baseline for later windows such as 2026–2032, for comparison with other fields, and as guidance for editors, funders and students.

Interesting ideas

• I1. A single normalized number divided by (N − 1) summarizes how evenly items fall into categories, so distributions with different category counts and corpus sizes can be compared. — For Mova (learn): Compute such a dispersion score over deprels, frame types or error classes in each silver batch, so we can spot when a corpus or a failure profile is collapsing onto a few categories.
• I2. Reliability is checked by double-coding a stratified random slice and reporting Cohen's κ, rather than trusting a single annotator. — For Mova (learn): Apply κ on a stratified sample to agreement between witness parsers or LLM teacher tiers. This gives a quantitative gate before labels enter silver data.
• I3. The coding scheme uses mutually exclusive categories with short operational definitions, plus an explicit catch-all for items that do not fit. — For Mova (frames): Speech-act or frame inventories could follow this pattern, giving every label a one-line operational test and a visible "other" bin. Tracking that bin shows where the inventory needs extending.
• I4. Tied frequencies get the average of the ranks they span, so ordering stays well defined when counts are equal. — For Mova (answer): Rank-based scoring of candidate answers or readings could use averaged ranks for ties. This avoids arbitrary tie-breaking that is hard to explain.
• I5. The paper's own explanation of which extreme gives Δ = 0 or 1 does not match its stated ranking rule, which shows the value of testing a metric on boundary cases. — For Mova (learn): Every new evaluation metric should ship with tiny synthetic extreme cases (all in one class, perfectly uniform) as negative controls. These confirm that the metric moves in the intended direction.