Unsupervised word embeddings capture latent knowledge from materials science literature.

Area of Science:

• Materials Science
• Natural Language Processing
• Computational Chemistry

Background:

• Scientific knowledge is predominantly in text, posing analysis challenges for traditional and machine learning methods.
• Structured property databases offer machine-interpretable data but cover a limited portion of scientific literature.
• Supervised natural language processing (NLP) for information retrieval requires extensive hand-labeled datasets.

Purpose of the Study:

• To develop an unsupervised method for encoding materials science knowledge from scientific literature.
• To demonstrate the ability of this method to capture complex scientific concepts without human supervision.
• To show the potential for predicting future materials discoveries from existing publications.

Main Methods:

• Utilized information-dense word embeddings (vector representations of words) to encode scientific text.
• Applied an unsupervised machine learning approach, eliminating the need for human labeling.
• Trained embeddings on a large corpus of scientific literature.

Main Results:

• Word embeddings successfully captured implicit materials science knowledge, including periodic table structure and structure-property relationships.
• The unsupervised method predicted functional materials for applications years before their actual discovery.
• Demonstrated that latent knowledge for future discoveries is embedded within past publications.

Conclusions:

• Unsupervised word embeddings offer an efficient way to extract and utilize knowledge from scientific literature.
• This approach enables collective knowledge extraction and suggests a generalized method for scientific literature mining.
• The findings open avenues for discovering new materials and accelerating scientific research through automated knowledge extraction.