Materials information extraction via automatically generated corpus

Area of Science:

• Materials Science
• Natural Language Processing
• Machine Learning

Background:

• Information Extraction (IE) from unstructured text is crucial for computer understanding of natural language.
• Machine learning-based IE requires large, accurately labeled datasets, which are difficult and time-consuming to create in materials science.
• Manual labeling of materials data for IE is laborious and requires expert input.

Purpose of the Study:

• To develop a semi-supervised IE framework for the materials domain.
• To reduce manual intervention in creating labeled corpora for IE.
• To enable automatic generation of materials corpora for IE tasks.

Main Methods:

• A semi-supervised IE framework utilizing an automatically generated corpus.
• Application of Snorkel for automatic labeling of material property values within the corpus.
• Training an IE model using an Ordered Neurons-Long Short-Term Memory (ON-LSTM) network on the generated corpus.

Main Results:

• Achieved F1-scores of 83.90% for γ' solvus temperature, 94.02% for density, and 89.27% for solidus temperature in superalloy data extraction.
• Demonstrated the framework's universality through successful application to other materials.
• Significant reduction in manual effort for corpus generation.

Conclusions:

• The proposed semi-supervised IE framework effectively generates labeled corpora for materials science.
• The framework demonstrates high accuracy and universality across different material types.
• This approach offers a scalable solution for advancing IE in materials informatics.