Task 01
n=484+10.34 ppperfect correction
V3.095.59%
V2.085.25%
← Yanlan product overview
Product upgrade · Yanlan V3.0
A new standard for Chinese pre-publication correction.
V3.0 finds more of the errors that matter, reduces unnecessary alerts, and returns a publication-ready correction more consistently across news, publishing, government, and regulated-content workflows.
95.74%Accuracy
7.04%Missed-error rate · FPR
1.71%False-alarm rate · FNR
94.16%Perfect correction rate
01 · Evidence
V3.0 compared with V2.0, at a readable scale.
Four deployment-critical metrics retain paired absolute values; the complete eight-metric ledger below shows the full sample-weighted result. V3.0 improves correction quality while lowering both missed errors and false alarms.
AccuracyHigher is better
V2.092.42%
V3.095.74%
+3.32 pp
Missed-error rate · FPRLower is better
V2.013.06%
V3.07.04%
−6.02 pp
False-alarm rate · FNRLower is better
V2.01.99%
V3.01.71%
−0.28 pp
Perfect correction rateHigher is better
V2.084.66%
V3.094.16%
+9.50 pp
Detection & alert control4 metrics
Accuracy+3.32 pp
Missed-error rate · FPR ↓−6.02 pp
False-alarm rate · FNR ↓−0.28 pp
Correction recall · C-Recall+6.02 pp
Correction quality4 metrics
Correction precision · C-Precision+0.52 pp
F1+2.84 pp
F0.5+4.39 pp
Perfect correction rate+9.50 pp
Evaluation protocol: V2.0 and V3.0 use the same task definitions; 4,408 examples across three high-difficulty sample sets (n=484 / 991 / 2933), weighted by sample count. Lower is better for FPR and FNR; higher is better for the other six metrics.
02 · Cross-task consistency
The largest gain holds across all three tasks.
Perfect correction exceeds 93% on every evaluation task. The improvement over V2.0 remains between +9.10 and +10.60 percentage points rather than depending on one favorable sample set.
Task 02
n=991+10.60 ppperfect correction
V3.094.82%
V2.084.22%
Task 03
n=2,933+9.10 ppperfect correction
V3.093.78%
V2.084.68%
Total examples4,408
Sample sets3
Direct comparisons47
Aggregate weightingBy sample count
03 · The upgrade
Built around the point where content becomes public.
V3.0 focuses on high-consequence pre-publication text: subtle wording errors, context-dependent corrections, dense domain vocabulary, and workflows where a plausible but unnecessary edit is itself a production risk.
A harder evaluation standard
The evaluation set was rebuilt around difficult, context-heavy examples drawn from high-standard content workflows instead of routine typo detection.
More corrections that can ship
Perfect correction rises by 9.50 percentage points, strengthening the model's ability to return a fix an editor can adopt directly.
Tighter alert control
Missed errors fall sharply while false alarms also decline, preserving editorial attention and supporting stable integration into production review.
04 · System-level result
Not one favorable metric. A consistent upgrade.
Across every dataset and metric tracked, V3.0 wins 43 of 47 direct comparisons. All eight sample-weighted aggregate metrics improve over V2.0.
43/47direct comparisons won
91.5% of measured comparisons favor V3.0.
The result spans detection, correction, precision, recall, and final correction quality—evidence of a system-wide improvement rather than a single benchmark spike.
Single-task × metric35–4
V3.0 leads 35 of the 39 comparisons measured within the three individual sample sets.
Sample-weighted aggregate8–0
After weighting by sample count, V3.0 improves every one of the eight aggregate metrics.
05 · Deployment
For high-standard content operations.
Yanlan V3.0 is now fully available for enterprise and government deployments, including private environments and workflows with organization-specific terminology.
News & publishingPre-publication copy, books, journals, and final editorial review.
Government communicationsOfficial documents, public notices, and policy communication.
Financial & legalCompliance-sensitive documents and regulated-content review.
Broadcast mediaSubtitles, transcripts, OCR text, and release-quality control.
Yanlan V3.0 is now fully available.
Enterprise and government teams can evaluate the upgrade against representative production text and deployment requirements.