WBA: Word Boundary Attention for Chinese Named Entity Recognition

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Chinese language exhibits unique structural properties with parallel word structures and sequential character connections.
• Existing methods may not fully capture the nuanced relationships between characters and words in Chinese text.

Purpose of the Study:

• To develop a novel attention mechanism for Chinese natural language processing that explicitly models word boundaries.
• To enhance character representations by incorporating contextual information derived from word lattices.
• To improve word recognition and leverage lexicon-based context in sequence labeling tasks.

Main Methods:

• Extraction of head and tail character positions within words for a relative positional encoding scheme.
• Introduction of Word Boundary Attention (WBA) to assign dynamic attention weights to characters.
• Enhancement of character representations using contextual information from the word lattice.

Main Results:

• WBA consistently outperforms existing approaches across multiple datasets.
• A 2.51% improvement over the base model was achieved on the Weibo dataset with YJ lexicon encoding.
• Visualizations revealed interpretable relationships between words and characters, offering insights into word discovery.

Conclusions:

• Word Boundary Attention effectively models word boundaries, suppressing noise and improving word recognition in Chinese NLP.
• The proposed method leverages richer lexicon-based context, leading to superior performance.
• The approach provides interpretable insights into the interaction between word-level and character-level information.