Enhancing Chinese Character Representation With Lattice-Aligned Attention

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Deep Learning

Background:

• Word-character lattice models are effective for Chinese NLP tasks by integrating word boundary information.
• Existing models struggle to capture fine-grained correlations within word-character spaces due to unidirectional processing.

Purpose of the Study:

• To propose a novel lattice-aligned attention network (LAN) for enhancing character representations in Chinese NLP.
• To model dense interactions within the word-character lattice structure.

Main Methods:

• Developed a lattice-aligned attention network (LAN).
• Incorporated a cross-lattice module, a gated word-character semantic fusion unit, and a self-lattice attention module.
• Focused on capturing explicit, fine-grained correlations across word-to-character and character-to-character spaces.

Main Results:

• LAN significantly improves model performance on Chinese NLP tasks.
• Achieved state-of-the-art results on three benchmark datasets.
• Demonstrated superior performance compared to existing competitive approaches.

Conclusions:

• The proposed LAN effectively models dense interactions in word-character lattices.
• LAN enhances character representations by capturing fine-grained correlations, leading to improved NLP task performance.
• LAN represents a significant advancement in Chinese NLP.