Handwriting recognition boosted by tokenization and data augmentation

Researchers have investigated two strategies for improving IMU-based online handwriting recognition: subword tokenization and data augmentation. Experiments on the OnHW-Words500 dataset showed that Bigram tokenization effectively reduces word error rates by improving generalization to unseen writing styles. However, tokenization proved detrimental for writer-dependent tasks. In contrast, concatenation-based data augmentation significantly reduced character and word error rates, outperforming extended training and effectively addressing intra-writer distributional sparsity. AI

IMPACT This research offers methods to improve the accuracy and generalization of handwriting recognition systems, potentially impacting applications requiring real-time input interpretation.

RANK_REASON Academic paper detailing a systematic study of methods for improving a specific AI task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• bigram
• Jindong Li
• OnHW-Words500

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →