Application of machine reading comprehension techniques for named entity recognition in materials science

Area of Science:

• Materials Science
• Computational Chemistry
• Information Retrieval

Background:

• Manual analysis of materials science literature is time-consuming.
• Traditional named entity recognition (NER) methods struggle with semantic information and nested entities.
• Automated extraction of material entities is crucial for knowledge graph construction.

Purpose of the Study:

• To develop an improved method for named entity recognition (NER) in materials science.
• To address limitations of sequence labeling methods in capturing semantic information and nested entities.
• To enhance knowledge extraction and data analysis in materials science literature.

Main Methods:

• Converted the traditional sequence labeling NER task into a machine reading comprehension (MRC) task.
• Utilized MRC to effectively extract multiple overlapping and nested entities.
• Integrated prior knowledge through queries for a deeper contextual understanding.

Main Results:

• Achieved state-of-the-art (SOTA) performance on multiple benchmark datasets (Matscholar, BC4CHEMD, NLMChem, SOFC, SOFC-Slot).
• Reported high F1-scores: 89.64% (Matscholar), 94.30% (BC4CHEMD), 85.89% (NLMChem), 85.95% (SOFC), and 71.73% (SOFC-Slot).
• Demonstrated effective utilization of semantic information and extraction of nested entities.

Conclusions:

• The proposed MRC-based NER method significantly enhances entity extraction accuracy and efficiency in materials science.
• This approach provides robust support for building knowledge graphs and advancing data analysis in the field.
• The innovation accelerates the discovery and development of new materials through improved information processing.