search

materialsintelligence / matscholar

Code to access the Matscholar public API.
https://matscholar.com
MIT License
61 stars 19 forks source link

NER can be quite off sometimes #52

Open computron opened 5 years ago

computron commented 5 years ago

e.g. the abstract for https://www.researchgate.net/publication/273349137_Realization_of_narrowband_thermal_emission_with_optical_nanostructures

seems to have a lot of inaccuracies

computron commented 5 years ago

another example: https://link.springer.com/article/10.1007/s11708-009-0009-x

LeighWeston86 commented 5 years ago

Looks like most of these errors are related to the application tag, which is one of our worst performing tags. There aren't a lot of papers like these in our NER training set.

jppgks commented 5 years ago

@LeighWeston86 is the NER training set available for download somewhere?

computron commented 5 years ago

Yes the training set is here: https://figshare.com/articles/Materials_Science_Named_Entity_Recognition_train_development_test_sets/8184428

jppgks commented 5 years ago

Thanks @computron!

littleflow3r commented 4 years ago

@LeighWeston86 is it possible to share the code used for training? i tried to reproduce the result but i only got like 76% accuracy, while in the paper it is mentioned abt 87%. Thanks!