Machine-Guided Polymer Knowledge Extraction Using Natural Language Processing: The Example of Named Entity...

Area of Science:

• Materials Science
• Computational Linguistics
• Natural Language Processing

Background:

• Extracting structured information from materials science literature using named entity recognition (NER) is an active research area.
• A significant challenge remains in normalizing these extracted entities, i.e., identifying when different textual mentions refer to the same real-world entity.
• Polymers, in particular, often possess multiple common and IUPAC names, complicating normalization efforts.

Purpose of the Study:

• To develop and evaluate a method for normalizing polymer named entities extracted from scientific literature.
• To address the challenge of recognizing that different names refer to the same polymer material.
• To improve the accuracy and efficiency of knowledge extraction pipelines in polymer science.

Main Methods:

• Trained supervised clustering models using pre-existing Word2Vec and fastText word embeddings.
• Utilized parameterized cosine distance functions for clustering and normalizing textually derived polymer entities.
• Employed a labeled dataset of polymer names for model training and validation.

Main Results:

• Achieved an F1 score of 0.85 in normalizing polymer named entities.
• Identified 6734 unique polymer clusters (unique polymers) from approximately 15,500 named entities in a corpus of 0.5 million papers.
• Manually curated 632 unique polymer clusters to train the normalization model.

Conclusions:

• The developed method effectively normalizes polymer named entities, significantly advancing the field of automated knowledge extraction.
• This work provides a critical component for natural language processing pipelines aimed at efficiently processing and understanding the vast polymer literature.
• The findings contribute to a more accurate representation of unique polymers discussed in scientific publications.