An ontology-based text mining dataset for extraction of process-structure-property entities.

Area of Science:

• Materials Science
• Computational Linguistics
• Knowledge Representation

Background:

• Large language models (LLMs) excel at statistical language representation.
• Ontologies offer symbolic knowledge representation, complementing LLMs.
• Neurosymbolic models require integrated text and ontology datasets for training and benchmarking.

Purpose of the Study:

• Introduce the MaterioMiner dataset and its associated materials mechanics ontology.
• Facilitate the training and benchmarking of neurosymbolic models at the intersection of LLMs and ontologies.
• Enable research in materials language models, automated ontology construction, and knowledge graph generation.

Main Methods:

• Developed a dataset intertwining a text corpus with a materials mechanics ontology.
• Annotated 2191 entities across four publications with 179 distinct classes, ensuring fine-grained detail.
• Explored inter-rater annotation consistency and fine-tuned pre-trained language models for named entity recognition.

Main Results:

• Established the MaterioMiner dataset with a linked materials mechanics ontology.
• Demonstrated high annotation consistency among three raters.
• Showcased the feasibility of training named entity recognition models using the dataset.

Conclusions:

• The MaterioMiner dataset is a valuable resource for advancing neurosymbolic AI in materials science.
• It supports the development of materials language models and automated knowledge extraction.
• Facilitates the creation of knowledge graphs from scientific literature.