Ad-hoc Personalization of Offline Handwritten Text Recognition Using Style Transfer

ACL ARR 2025 May Submission62 Authors

Abstract: Personalizing handwritten text recognition can significantly enhance recognition accuracy, but requires a substantial number of handwritten samples, which limits its practical relevance. In this work, we investigate the potential of style-transferred synthetic samples for ad-hoc personalization. We show that one-shot and few-shot generators are able to produce visually similar handwriting samples. However, our experiments also show that the style-transferred data has no measurable personalization effect. This finding holds in a fair comparison with the same amount of samples and when using larger quantities of synthetic data.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: multimodal applications, data augmentation

Contribution Types: Model analysis & interpretability, Reproduction study, Approaches to low-resource settings

Languages Studied: English

Submission Number: 62