Run COGNILANG in your languages
The protocol was built so that a team working on Spanish–Chinese, German–Turkish, Swahili–English, Hindi–Urdu — any pair or triad with substantial cross-language news flow — can produce results commensurable with the EN/FR/AR pilot. The rule is simple: the scientific core is fixed; the language layer is adapted and documented.
What stays fixed · what you adapt
Invariant core — do not modify
· The constructs and hypotheses H1–H6 (your language pair substitutes into them; their logic is unchanged)
· The four code families LIN DIS
TRA PRD and their definitions
· Corpus structure: comparable corpus + parallel sub-corpus + hand-annotated gold subset
· Validation rules: κ ≥ 0.70 on doubled annotation; every AI measure benchmarked on the gold subset;
prompts and model versions logged
· The three-layer causal discipline: language system vs. translational mediation vs. editorial context
· Reader-study constructs: comprehension, recall, cognitive load, credibility, cloze
· Epistemic rule: no cognitive conclusions from text alone
Adaptation layer — adapt and report
· Tokenisation, segmentation, and orthographic normalisation for your languages (tools and settings documented, versions logged)
· The DIS-1 framing label set: start from the pilot set, pilot it on 10 articles in your languages, refine once, freeze, report changes
· One monolingual language model per language for PRD-1 surprisal, of comparable scale where possible
· Outlet selection and the comparability window (e.g. ±72 h), justified for your media landscape
· Reader-study instruments translated and back-checked; consent forms in all study languages
· Script-specific handling: RTL rendering, diacritics policy, romanisation conventions, CJK segmentation
Eight steps from registration to results
Register your language pair
Apply via the Network & Join page with your pair/triad, affiliation, and corpus access. Registration prevents duplicate pairs and connects you to the coordination team and the shared instruments.
Fix your scientific architecture (Week 1)
Instantiate RQ0–RQ6 and H1–H6 for your pair; write operational definitions (framing, cognitive load, linguistic prediction, the shift types); initiate your institution's research-ethics application immediately.
Build the literature bridge (Week 2)
Review the shared bibliography, then add the literature specific to your languages: news translation in your media sphere, psycholinguistics of your scripts, available NLP resources.
Design and collect the corpus (Week 3)
Same events covered in all your languages, within your comparability window; 100–300 articles per language; ≥ 30 source/translation pairs per direction; full provenance log.
Adapt, pilot, and freeze the grid (Week 3)
Translate the code-book; add one positive and one negative example per code in your languages; pilot on 10 articles; double-annotate ≥ 15 % of the gold subset; reach κ ≥ 0.70; freeze.
Stand up the AI pipeline (Week 3)
Configure language tools, alignment models, and one monolingual LM per language for surprisal; run end-to-end on a 10-article sample; log every issue and every version.
Analyse and cross-link (Week 4)
Comparative LIN/DIS profiles per event; TRA typology with frequencies; predictability profiles per version; the shift-type × surprisal-change cross-table; results and discussion against H1–H6.
Report back to the network (Weeks 5–6)
Deliver the standard report plus a structured deviation log (everything adapted, and why), the frozen code-book, and your gold-subset annotations — the currency of comparability.
Known adaptation points by language type
| If your languages include… | Pay particular attention to |
|---|---|
| Right-to-left scripts (Arabic, Hebrew, Persian, Urdu) | Orthographic normalisation (letter variants, diacritics policy); RTL rendering in every output format; bidi handling in mixed-script examples. |
| Morphologically rich languages (Arabic, Turkish, Finnish, Russian) | Tokenisation choices materially change every count — document them; surprisal concentrates differently across morphemes; prefer per-idea-unit aggregation. |
| Unsegmented scripts (Chinese, Japanese, Thai) | Word segmentation tool and settings are part of the method; report segmentation agreement on a sample; align at clause level if word alignment is unstable. |
| Closely related pairs (Spanish–Portuguese, Hindi–Urdu) | Translation vs. light editing is harder to distinguish — tighten the criteria for the parallel sub-corpus and code provenance conservatively. |
| Low-resource languages | Model quality limits PRD-1: report model perplexity on held-out news; lean more on PRD-2 (human cloze) and PRD-3 (device counts); flag tool limitations explicitly in methods. |
| Diglossic situations (Arabic varieties, Swiss German) | Fix the register of the corpus (e.g. Modern Standard Arabic) and state it; treat register shifts toward/away from the standard as TRA-7 with a note. |
Frequent questions
Do we need three languages?
No. The minimum is one pair with a real translation flow between them. Triads add a comparative panel but are not required for commensurability.
Can we skip the reader study?
You can defer execution, but you must design it: the approved-ready protocol, instruments, and pilot materials are a standard deliverable. Text-only replications report hypotheses about reception, not effects.
Must we use the same AI models?
No — you must use the same validation logic: whatever models you choose, benchmark them on your gold subset and log versions, prompts, and parameters. Model substitution is expected across languages.
Who owns our data and results?
Your team. The network asks for the deviation log, frozen code-book, gold-subset annotations, and derived measures — not raw copyrighted texts — under the data policy on the Join page.
How long does a replication take?
The reference format is a six-week intensive workshop (the pilot's plan, available in the worksheet pack). A part-time semester works equally well; the sequence matters more than the calendar.
What if our framing labels don't fit?
Expected. Pilot the DIS-1 set on 10 articles, adapt once, freeze, and report the mapping between your labels and the pilot set so cross-team comparison stays possible.