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

Abstract

This repository contains named-entity recognition (NER) datasets for four materials science and engineering (MSE) publications and utility functions to handle the data. The scope of the scholarly articles used as a basis is crystallographic defects, microstructure, mechanical properties in particular fatigue. Each annotation corresponds to a class in a materials science domain ontology called materials mechanics ontology. This should prospectively enable linking materials knowledge and data to facilitate training neurosymbolic machine learning models. Two dataset variants are published:

coarse-granular named-entity recognition (CG-NER) where the annotated concepts are high-level ontological classes fine-granular named-entity recognition (FG-NER) where the annotated concepts are low-level ontological classes

Aside from the link to the ontology a characteristic of the dataset is its high degree of of annotation. Namely, 179 distinct ontological classes and 27% of all tokens are annotated in the fine-granular dataset.