Table 5.

Cross validation results on the training data set (5-fold cross validation) for subtask 1 using the attention-based bidirectional long short-term memory network coupled with conditional random fields with different types of embeddings. When using word embeddings, some configurations enabled embedding fine-tuning for 2 epochs. For simplicity purposes, only F1 scores are presented.

Embeddings type and model configuration		Family member	Observations	Overall
clinicalBERTa
	Baseline	0.4103	0.8596	0.7194
	Baseline+EDb	0.3788	0.8481	0.7023
	Baseline+Neji	0.3545	0.8478	0.6908
	Baseline+ED+Neji	0.3485	0.8688	0.7081
BioWordVec
	Baseline	0.5921	0.8140	0.7317
	Baseline+ED	0.6553	0.8276	0.7627
	Baseline+ETc	0.6166	0.8285	0.7513
	Baseline+ED+ET	0.6219	0.8367	0.7579
	Baseline+ED+Neji	0.7222	0.8529	0.8036
	Baseline+ED+ET+Neji	0.7266	0.8587	0.8092

^aclinicalBERT: clinical bidirectional encoder representations from transformers.

^bED: entity discovery.

^cET: embeddings training.