Talk to Me

OCR Text Classification Model Comparison

Binary Classification Performance Analysis - Form Type Recognition

Dataset: 9,990 samplesClasses: mednarr (779) | non-mednarr (9211)

Key Findings

Winner Based on False Positives

Longformer-DeBERTa has fewer false positives

False Positives: DistilBERT: 60 | Longformer: 28

Standout Metric

DistilBERT achieves 99% recall on mednarr class (only 6 false negatives)

Trade-off Analysis

Longformer: Better precision (96% vs 93%) but lower recall (95% vs 99%)

Error Pattern

DistilBERT: More false positives (60 vs 28). Longformer: More false negatives (42 vs 6)

Performance Radar