Towards structuring clinical texts: joint entity and relation extraction from Japanese case report corpus

Setpu

BERTs and dataset can be downloaded from the following.

An NVIDIA RTX A6000 (48GB) was used for training the models. Modify the arguments in the code according to your environment.

bash run_joint.sh

Fold	NER (UTH)	RE (UTH: Strict)	NER (NICT)	RE (NICT)
1	0.911	0.761	0.894	0.745
2	0.909	0.757	0.886	0.735
3	0.910	0.759	0.890	0.740
4	0.918	0.759	0.901	0.747
5	0.910	0.758	0.887	0.736
Avg.	0.912	0.759	0.892	0.741

CC BY-NC-SA 4.0

Ma, Y., Hiraoka, T., & Okazaki, N. (2022). Named entity recognition and relation extraction using enhanced table filling by contextualized representations. Journal of Natural Language Processing, 29(1), 187-223.. Software available from https://github.com/YoumiMa/TablERT.
pytorch-crf. Software available from https://pytorch-crf.readthedocs.io/en/stable/.
Hiroki Nakayama. seqeval: A python framework for sequence labeling evaluation, 2018. Software available from https://github.com/chakki-works/seqeval.