Named Entity Recognition and Normalization Applied to Large-Scale Information Extraction from the Materials Science Literature

Abstract

Over the past decades, the number of published materials science articles has increased manyfold. Now, a major bottleneck in the materials discovery pipeline arises in connecting new results with the previously established literature. A potential solution to this problem is to map the unstructured raw-text of published articles onto a structured database entry that allows for programmatic querying. To this end, we apply text-mining with named entity recognition (NER), along with entity normalization, for large-scale information extraction from the published materials science literature. The NER is based on supervised machine learning with a recurrent neural network architecture, and the model is trained to extract summary-level information from materials science documents, including: inorganic material mentions, sample descriptors, phase labels, material properties and applications, as well as any synthesis and characterization methods used. Our classifer, with an overall accuracy (f1) of 87% on a test set, is applied to information extraction from 3.27 million materials science abstracts-the most information-dense section of published articles.Overall, we extract more than 80 million materials-science-related named entities, and the content of each abstract is represented as a database entry in a structured format. Our database shows far greater recall in document retrieval when compared to traditional text-based searches due to an entity normalization procedure that recognizes synonyms. We demonstrate that simple database queries can be used to answer complex \meta-questions" of the published literature that would have previously required laborious, manual literature searches to answer. All of our data has been made freely available for bulk download; we have also made a public facing application programming interface (https://github.com/materialsintelligence/matscholar) and website http://matscholar.herokuapp.com/search for easy interfacing with the data, trained models and functionality described in this paper. These results will allow researchers to access targeted information on a scale and with a speed that has not been previously available, and can be expected to accelerate the pace of future materials science discovery.

Automated extraction of chemical synthesis actions from experimental procedures

“…It does not focus solely on experimental procedures and is also able to extract spectroscopic attributes or information present in tables, for instance. Weston et al follow a similar strategy and apply their method on materials science abstracts with the goal to produce easily searchable knowledge databases 18 .…”

Computational discovery of promising new n-type dopable ABX Zintl thermoelectric materials

“…These newly predicted compounds comprise 13 phosphides, 17 arsenides, 8 antimonides, and 7 bismuthides ( Figure 5). A survey of the published literature using Matscholar (www.matscholar.com) 48 further confirms that these 45 compounds are indeed new and previously unreported (except LiCdSb 45 ). The larger number of undiscovered arsenides could be due to synthesis challenges associated with handling As, which is toxic.…”

Can we predict materials that can be synthesised?

“…Unsupervised and semi-supervised machine learning have been recently used to by-pass that problem, allowing one to reconstruct not only owcharts of possible synthetic procedures, but also to capture latent chemical knowledge such as the prediction of promising material candidates years prior their publication. [14][15][16] However, the quality of the datasets and the predictive power of literature-based models can only be as good as the original literature data. Many technical details and conditions are oen omitted from reported procedures, which rely on the experience of the experimentalist and hidden instrument settings.…”