Self-Supervised Discovery of Cross-Lingual Shared Knowledge for Continual Text Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Incremental multilingual text recognition (IMLTR) faces challenges with cross-lingual knowledge sharing.
• Current IMLTR methods assume single-language domains, neglecting inherent multilingual data properties.
• The incremental sharing problem, where cross-lingual knowledge is overlooked, hinders continual learning.

Purpose of the Study:

• To propose a novel self-supervised cross-lingual knowledge discovery framework, CrossKnow, for IMLTR.
• To address the limitations of existing IMLTR methods in handling multilingual scenarios.
• To improve continual learning in IMLTR by mitigating knowledge forgetting and enabling efficient knowledge transfer.

Main Methods:

• Developed a shared knowledge discovery strategy using prediction consistency across recognizers.
• Eliminated reliance on language labels for individual characters.
• Designed a multi-granularity, multi-task language domain discriminator to capture inter-language dependencies.
• Guided hierarchical sequence decoding using discovered shared knowledge.

Main Results:

• CrossKnow effectively mitigates the forgetting of previously learned knowledge.
• The framework achieves efficient cross-lingual knowledge transfer.
• Experiments on MLT17 and MLT19 datasets show CrossKnow's superiority.
• Achieved competitive performance without additional language supervision, reducing storage and improving efficiency.

Conclusions:

• CrossKnow significantly advances incremental multilingual text recognition.
• The proposed method offers a more accurate and efficient approach to continual learning in multilingual contexts.
• CrossKnow demonstrates the effectiveness of self-supervised cross-lingual knowledge discovery for IMLTR.