Unsupervised word embeddings capture latent knowledge from materials science literature

A Critical Review of Machine Learning of Energy Materials

“…Furthermore, the vast majority of databases are commercial products requiring a license, and programmatic application programming interfaces (APIs) for large‐scale data access are rarely implemented. A large fraction of experimental data are only available in journal publications, though recent successes in text mining offer a potential solution to this conundrum . Finally, major efforts are underway in high‐throughput/combinatorial experiments that can generate large experimental materials database with diverse properties …”

“…While PCA works on linear projection, the manifold learning is able to capture nonlinear relationships. For example, the t‐distributed stochastic neighbor embedding (t‐SNE) method learns low‐dimensional representations such that the local distance between data points is roughly preserved and has been applied in visualizing the elemental embedding vector trained from materials property prediction models, structural similarity of perovskites, word embeddings in text mining, electronic fingerprints, etc.…”

“…Encouraged by the advancement in the natural language processing field, attempts have been made to extract data and insights from literature using text mining. In thermoelectrics, the relevant literature consists of 3.3 million published papers between 1922 and 2018 in more than 1000 journals . Tshitoyan et al has shown that, by feeding these abstracts to a “machine,” one can obtain embeddings of words that incorporate the chemical relationships.…”

“…Examination of the context words indicates that the algorithm makes decision based on structure type, co‐mention of other thermoelectric materials, association of other related applications, and description of the materials' properties. Reproduced with permission . Copyright 2019 Nature Publishing Group.…”

“…A large fraction of experimental data are only available in journal publications, though recent successes in text mining offer a potential solution to this conundrum. [48][49][50][51][52][53] Finally, major efforts are underway in high-throughput/combinatorial experiments that can generate large experimental materials database with diverse properties. [31] It is therefore not surprising that many ML works have turned to computed data sources.…”