Sequential lexicon enhanced bidirectional encoder representations from transformers: Chinese named entity recognition...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Lexicon Enhanced Bidirectional Encoder Representations from Transformers (LEBERT) has shown success in Chinese Named Entity Recognition (NER).
• LEBERT's Lexicon Adapter layer fuses lexicon knowledge but may introduce noise and fail to address word conflicts.
• Existing methods lack robust mechanisms for handling lexicon noise and inter-word conflicts.

Purpose of the Study:

• To introduce a novel lexical enhancement method, Sequential Lexicon Enhanced BERT (SLEBERT), for Chinese NER.
• To address the limitations of LEBERT, specifically noise word introduction and lexical conflicts.
• To improve the performance and efficiency of Chinese NER through enhanced lexicon integration.

Main Methods:

• Developed SLEBERT, a model that builds a sequential lexicon to mitigate noise and resolve conflicts.
• Incorporated position encoding of the sequential lexicon to enhance feature representation.
• Utilized an adaptive attention mechanism within the sequential lexicon for improved feature fusion.

Main Results:

• SLEBERT demonstrated superior performance compared to existing lexical enhancement models on four Chinese NER datasets.
• The proposed sequential lexicon approach effectively reduced noise words and resolved lexical conflicts.
• SLEBERT achieved higher efficiency in processing and feature extraction.

Conclusions:

• SLEBERT offers a more effective and efficient approach to lexical enhancement for Chinese NER.
• The sequential lexicon strategy successfully overcomes limitations associated with noise and conflict in previous methods.
• SLEBERT represents a significant advancement in applying lexicon knowledge to deep learning models for NER.