Dynamic meta-learning with generative augmentation for cross-lingual Japanese few-shot named entity recognition

Named Entity Recognition (NER) in Japanese is a challenging task due to data scarcity, limited cross-lingual transfer capabilities, and fuzzy entity boundaries, especially in low-resource environments. This research presents a novel framework, MAML-ProtoNet++, designed to overcome these challenges. The framework combines Model-Agnostic Meta-Learning (MAML), which allows for rapid parameter adaptation, with Prototypical Networks (ProtoNet) that perform prototype-based classification for few-shot learning. Additionally, the framework integrates cross-lingual contrastive pretraining using the multilingual mT5 model, which generates diverse pseudo-samples and optimizes the semantic alignment between Japanese and English entity pairs. To address the problem of insufficient annotated data, generative augmentation techniques and boundary verification methods are employed, improving the support set and entity boundary recognition. The experimental results demonstrate that MAML-ProtoNet++ outperforms existing models with a macro-average F1 score of 0.772 under a 5-shot setting. The boundary recognition accuracy is notably high, with 0.85 for start points and 0.84 for end points. Additionally, cross-lingual pretraining significantly improves semantic alignment, with cosine similarity between Japanese and English entities increasing from 0.61 to 0.85. These results highlight the robustness and adaptability of MAML-ProtoNet++ in handling complex few-shot and cross-lingual NER tasks. The findings suggest that this framework is a promising solution for NER in low-resource languages like Japanese, offering potential for broader applications in cross-lingual transfer learning.