MIT Press

Table 4:

Test performance on the TACRED relation extraction benchmark. BERT_large and BERT_EM+MTB from Soares et al. (2019) are the current state-of-the-art. ^*: BERT_EM+MTB incorporated an intermediate “matching the blanks” pre-training on the entity-linked text based on English Wikipedia, which is not a direct comparison to ours trained only from raw text.

	p	R	F1
BERT_EM(Soares et al., 2019)	−	−	70.1
BERT_EM+MTB^*	−	−	71.5
Google BERT	69.1	63.9	66.4
Our BERT	67.8	67.2	67.5
Our BERT-1seq	72.4	67.9	70.1
SpanBERT	70.8	70.9	70.8

	p	R	F1
BERT_EM(Soares et al., 2019)	−	−	70.1
BERT_EM+MTB^*	−	−	71.5
Google BERT	69.1	63.9	66.4
Our BERT	67.8	67.2	67.5
Our BERT-1seq	72.4	67.9	70.1
SpanBERT	70.8	70.9	70.8

View Large

This Feature Is Available To Subscribers Only