Improving Zero-Shot Entity Linking With Chinese Feature Convolution Embeddings

Zero-shot entity linking (ZEL), a fundamental technique for many downstream natural language processing tasks, aims to accurately associate textual mentions with specific entities in a knowledge base. However, Chinese ZEL remains challenging due to the limited research on Chinese entity linking and emerging entities. The predominant two-stage model often overlooks the unique characteristics of Chinese, which limits its ability to capture a wider range of semantic information. To address this, we propose CFCE-ZEL, a ZEL model based on Chinese Feature Convolutional Embeddings (CFCE), which is built upon the Bidirectional Encoder Representations from Transformers (BERT) architecture for contextualized representations and integrates a Convolutional Neural Network (CNN) for Chinese feature extraction. The model adopts a two-step framework that generates and ranks candidate entities to produce the final entity linking results. It further employs a convolutional neural network and integrates Chinese-specific feature embeddings at both image and character levels, namely glyph and radical embeddings, to enhance word representations and mitigate the impact of out-of-vocabulary words. Moreover, we employ FAISS and knowledge distillation to improve the efficiency of the two-stage model while maintaining a balance between speed and performance. Extensive experiments on the Hansel dataset show that CFCE-ZEL outperforms multiple baselines in both candidate entity generation and ranking, while maintaining high efficiency with FAISS and knowledge distillation.