Application of Text Mining in Materials Science

Given the pervasive influence of text mining on various societal domains, the scientific domain in particular is witnessing an increasing emphasis on data acquisition. In the realm of materials science, an increasing demand exists for the extraction of knowledge, encompassing physical properties and synthesis processes of materials from a substantial volume of scientific papers. This field presents a unique challenge due to the inherent heterogeneity resulting from the different classes of materials studied and the range of material properties under investigation. As a result, the data generated spans a wide range of scales, in the form of numerical data, textual information or image-based content, requiring sophisticated quantitative interpretation. Using techniques borrowed from natural language processing, there is a compelling opportunity to automatically process and organise scientific literature across multiple fields, unlocking and generating the rich datasets essential for advances in data science and machine learning.

A review of advances and methodologies in natural language processing and text mining as applied to the materials science literature is presented in the work of Elsa A et al. [1]. They highlight the potential to extract valuable information beyond the textual content found in figures and tables within scholarly articles. This exploration is motivated by various objectives, including data collection, hypothesis development, and discerning trends within and across fields. The work delves into illustrative examples and outlines current and emerging natural language processing methods, highlighting their applications in the context of materials science. It also examines the challenges posed by natural language processing and the intricacies of data in materials science, and provides insights into possible directions for future developments.

Another example of the application of text mining methods, in this case in the field of inorganic materials science, is presented in the work of Kuniyoshi et al [2]. In this work, they introduce a large-scale Natural Language Processing (NLP) pipeline designed to extract material names and properties from the materials science literature to facilitate search and retrieval of results in the field. To achieve this goal, a label definition is proposed for the extraction of material names and properties, resulting in the creation of a corpus of 836 annotated paragraphs extracted from 301 papers. This corpus serves as training data for a named entity recognition (NER) model. Experimental results demonstrate the effectiveness of the NER model, which achieves successful extraction with a micro-F1 score of 78.1%. To further demonstrate the usefulness of the approach, a comprehensive evaluation is performed on a real-world automatically annotated corpus, using the trained NER model on 12,895 materials science papers.