Indonesian Sports Text Classification Modeling Based on Few-Shot NER

Abstract

This study investigates the development of a Few-Shot Named Entity Recognition (NER) model for Indonesian sports texts, addressing the persistent challenge of limited annotated resources. While conventional NER approaches such as SpaCy, BiLSTM-CRF, and transformer-based models typically require substantial training data, this work demonstrates that a Few-Shot learning strategy can achieve competitive and in many cases superior performance using only minimal labelled examples. By incorporating prototype-based adaptation and prompt-driven fine-tuning into a pre-trained SpaCy pipeline, the proposed model effectively distinguishes eight Indonesian sport categories despite severe token imbalance and low lexical diversity. This approach is particularly advantageous for this dataset, as Indonesian sports headlines are short, context-limited, and structurally compact, making Few-Shot methods better suited to capturing fine-grained entity distinctions than traditional high-data models. Experimental results show that the Few-Shot model outperforms all baselines, achieving a precision of 0.991 and an F1-score of 0.995, compared with the standard SpaCy model’s F1-score of 0.983. These improvements highlight the model’s enhanced generalization capability, reduced reliance on extensive annotation, and improved discrimination among closely related sport entities. The key contribution of this study lies in demonstrating the scalability and practical viability of Few-Shot NER for low-resource Indonesian NLP, offering an efficient and adaptable solution for domain-specific information extraction in scenarios where large annotated datasets are unavailable. The findings reinforce the potential of Few-Shot learning as a transformative method for sports analytics, media processing, and broader applications in under-resourced languages.