THE WORLD / JAPANESE-LANGUAGE EDUCATION
Japanese, beyond Japan.
Where a language is taught tells one story.
What learners read tells another.
SURVEY WAVES / 2006—2024
01 / THE ATLASCOUNTRY/AREA-LEVEL SURVEY DATA
Title-family evidence & sources
世界の日本語THE WORLD / GLOBAL DISTRIBUTION
Province / state directory
2024 public directory snapshot · distinct directory records, not survey totals or campuses. JF’s state/city field mixes administrative levels. Taiwan combines both directory groups. No local learner counts, textbook adoption or boundaries are inferred. Explore institutions at JF ↗
Height: square-root scale · Click a column or country · Scroll to zoom
Explore countries & areas
| COUNTRY / AREA | LEARNERS ↓ | SCRIPT BACKGROUND | TEXTBOOK COVERAGE |
|---|
03 / THROUGH TIME
The changing world of Japanese.
Compare countries, areas, regions and measures on one plot. Add selections below, then toggle individual series in the legend.
Reading this comparison
Hover or focus a point to inspect a value.
Year-by-year categories, sources and comparison limits
These are repeated surveys of institutions, not a fixed panel of the same schools or learners. Changes can reflect education activity, coverage and response. Japan is outside scope. Country names are linked across renamed entities; an absent country-year is missing, never an assumed zero. Geometry uses a single present-day base map, not historical borders.
Common stage view: Primary + secondary, higher education, and education outside schools. Native primary/secondary splits are retained where the source reports them. They are never estimated from a combined total. The multiple-stage category in 2009/2012 remains separate; stage-specific growth is not directly comparable with later redistributed learner counts.
Teacher counts: counting practices changed across waves, especially 2021→2024. The 2024 report counts a teacher in each educational stage taught within an institution, unlike the institution total collected in 2021. The teacher chart marks these breaks with dashed segments and suppresses growth percentages across a definition break.
Other survey categories: ownership, degree awards, teacher training, motivations, implementation problems/status and online teaching use different questionnaires or availability. This atlas links each year's original tables and documents their availability; it does not merge these categories into a fabricated comparable panel. Textbook mentions and cover previews remain a separate 2025 profile snapshot.
04 / COMPARATIVE STUDIES
Patterns across borders.
Compare education counts or ratios, learning settings and encoded textbook evidence. The saved default uses the strongest tested configuration on a documented cohort.
Exploratory groups, not validated learner types. Region and script background are inputs, so geographical separation is partly built in. Curation status may group countries by data availability; deselect it to check sensitivity.
How the mixed-data clustering works
Choose deterministic k-medoids or average-linkage clustering over a group-weighted mixed dissimilarity. Each selected block has equal weight; numeric counts and ratios use log(1+x) and within-cohort range scaling. A ratio is missing when its denominator is missing or zero. Ratio blocks and count blocks can be selected separately; using both can repeat scale-related information.
Textbook features: one binary presence per normalized title-family label per area, regardless of repeated stage records. Only reported-use evidence enters the encoding; historical/publication mentions and explicitly identified digital resources are excluded. Labels are Unicode/space normalized, not inferred ISBN matches. The vocabulary retains labels documented in at least two cohort areas. One-hot uses binary weights; TF–IDF uses binary TF × (log((1 + reviewed areas)/(1 + document frequency)) + 1), followed by L2 normalization. Cosine distance contributes one feature block, so a large vocabulary does not overwhelm numeric ratios. Empty or unreviewed vectors are missing, not assumed absence. Cluster cards show the strongest mean encoded features, not a mean title count.
Evaluated default: candidate feature sets, both encodings, both algorithms and k = 2…8 are compared on the same complete-case 2024 cohort. Selection balances silhouette and five deterministic 80% subsample stability checks (adjusted Rand index), subject to minimum-cluster-size and maximum-cluster-share safeguards. This is the best tested configuration under these criteria, not an objectively true partition. Region, script background and review status are excluded from default selection to avoid groups driven directly by those labels. They remain optional exploratory inputs. The evaluation and assignments are saved in the open data.
Missing groups are omitted pairwise and weights renormalized; a run with no comparable fields for a pair is rejected. The medoid minimizes within-group distance. Silhouette measures separation under the chosen distance, not pedagogical validity. Historical waves disable contemporary textbook features. Group numbers are recomputed per wave, not persistent identities.
Reference: mixed-type dissimilarities and missing values ↗05 / THE OPEN DATA DESK
Follow every record back to its source.
Download the evidence behind this edition: original title wording, PDF pages, education settings and review coverage. The survey panel spans 2006–2024; textbook evidence is a separate 2025 profile snapshot.
Material evidence · CSV ↓1,299 source-linked observationsMaterial evidence · JSON ↓Full nested scope and adoption fieldsSurvey panel · CSV ↓Seven waves · missing stays missingReview coverage · CSV ↓166 reviewed profiles · gaps disclosed
A publication mention is not adoption. A local percentage is not a national rate. Title families do not identify a specific edition or establish readability.