Introduction
Named Entity Recognition (NER) in Chinese presents unique challenges compared to alphabetic languages. The lack of explicit word boundaries and the ideographic nature of Chinese characters (Hanzi) mean that semantic clues are often embedded within the visual structure of the characters themselves. While previous "glyph-based" models have attempted to use either image-based CNN features or radical-based decompositions, they often fail to connect pictorial elements with abstract radicals.
Crucially, characters sharing identical radicals can have vastly different meanings based on their spatial arrangement. For example, the characters "叶" (leaf) and "古" (ancient) share the same components but arrange them differently. AHVE-CNER addresses this by explicitly aligning visual patches of a character image with its corresponding radical components.
Methodology: The AHVE-CNER Framework
AHVE-CNER introduces a multi-modal framework that explicitly aligns character images with their internal radical structures across three main stages:
1. Multi-Source Encoding
- Contextual Semantic: A pre-trained BERT model captures sentence-level context.
- Phonetic (Pinyin): A CNN processes character initials, finals, and tones to create pronunciation embeddings, providing a strong signal for disambiguation.
- Visual Alignment: Using a Vision Transformer (ViT), the model divides character images into patches and aligns them with radical embeddings from a dictionary. This grounded structural representation allows the model to distinguish between characters with similar components but different meanings.
2. Fusion via Gated Pangu-π Cross Transformer
Instead of simple concatenation, AHVE-CNER employs a fusion block inspired by the Pangu-π architecture. It utilizes Augmented Shortcuts and Series Activation Functions (SIAF) to capture complex inter-modal relationships. A Tri-modal Gating unit dynamically weights text, visual, and phonetic information for each character based on context.
3. Decoding
The framework uses W2NER to treat NER as a word-word relationship classification task, making it effective for both flat and nested entities.
Experimental Results
The model was evaluated on the Weibo (informal social media) and Resume (formal) benchmarks, setting new performance standards:
- Weibo: Achieved an F1-score of 73.09%, a 2.3% improvement over the previous state-of-the-art VisPhone.
- Resume: Reached an F1-score of 97.02%, outperforming other glyph-based models like FGN and MECT.
Ablation studies highlighted that while visual alignment is crucial for handling rare characters and spatial disambiguation, phonetic information remains a vital signal in modern Chinese NER.
Conclusion
AHVE-CNER demonstrates that aligning visual patches with radical components significantly enhances a model's ability to understand Hanzi structure. While the added complexity increases latency by approximately 27% compared to text-only models, the substantial gains in accuracy across diverse datasets establish a new state-of-the-art for Chinese NER. Future work will explore integrating word-level lexicons to further refine the model's performance.
Summary updated