|
|
HzTxt achieved accuracy comparable to BERT on clean text but maintained near-original performance on the adversarial TextFlipped dataset (where 20% of characters were randomly swapped). This confirms the hypothesis that spectral representation separates "semantic signal" from "orthographic noise."
In the West, the closest equivalent is the "Spline Font" used by early CNC machines, or the "Single Stroke" fonts on old HP plotters. But those were for letters. HZTXT had to solve for 6,763 common characters (GB2312).
HzTxt achieved accuracy comparable to BERT on clean text but maintained near-original performance on the adversarial TextFlipped dataset (where 20% of characters were randomly swapped). This confirms the hypothesis that spectral representation separates "semantic signal" from "orthographic noise."
In the West, the closest equivalent is the "Spline Font" used by early CNC machines, or the "Single Stroke" fonts on old HP plotters. But those were for letters. HZTXT had to solve for 6,763 common characters (GB2312). HzTxt achieved accuracy comparable to BERT on clean
