Progress in Natural Language Processing of Materials Science Text
- 5. 5
- 7. Scientific Domain Progression for NLP 7 Riedel and McCallum, Empirical methods in NLP 2011 ChemTagger: Hawizy et al 2011 ChemSpot: Rocktäschel et al 2012 Tchoua et al, IEEE, 2017 DOI 10.1109/eScience.2017.23 Biology/medical domain: Publicly available annotated collections GENIA corpus Unified Medical Language System Chemistry domain: Domain‐specific terminology, numeric phrases CHEMDNER corpus Inorganic materials: Polymer domain:
- 9. Continuum of text mining approaches • Collection of rules/dictionaries ; rule‐matching engine 9 J Chem Inf Model, 56, 2016 LeadMine: Lowe and Sayle, J Chem Inf Model, 7, 2015
- 10. Continuum of text mining approaches 10 Court and Cole., Scientific Data 2018 Korvigo et al., J Cheminform 2018CNN: convolutional neural net HS‐biGRU: half‐stateful bidirectional gated recurrent unit FCN: fully‐connected network CEM: chemical entity mention
- 13. Parts of a text, types of texts 13 Experimental methods NaNi1/3Co1/3Fe1/3O2 was synthesized by solid‐state reaction. Excess amounts of Na2O, NiO, Co3O4 and Fe2O3 were mixed and ball milled for 4 h at 500 rpm rate, and the resulting material was collected in the glove box. About 0.5 g of powder was fired at 800 °C under O2 for 14 h before it was quenched to room temperature and moved to a glove box filled with argon.
- 14. Challenges • With NLP for materials: • Transferability across materials domains • Off‐stoichiometry • Lexical ambiguity and evolution • With overall approach: • Age, quality • Linked recipes • Negative examples 14 “BaxMn1‐xO3 for x = 0.9” Solid state, thin film, templated synthesis “following the method described by….” Optical character recognition, pdf vs. html Bias of literature toward success
- 18. Some examples of what we’ve done… Variational autoencoder: • Loss = reconstruction + f(Gaussian) • Also a generative model Edward Kim et al., npj Computational Materials 2017 Collaborator, Stefanie Jegelka, CSAIL, MIT
- 20. Concluding thoughts on NLP progress • Natural language processing is young in its application to materials science • It takes effort to build up an annotation approach and corpus • There are domain‐specific needs regarding accuracy and ambiguity • Tradeoff between accuracy and degree of generalizability