MIT Press

Skip to Main Content

Table 5:

Results of IE embedding extrinsic evaluation via IE disambiguation—evaluated using F1 score (F1) and Accuracy (Acc%), and IE span detection—evaluated using sequence accuracy (Seq Acc%), and token-level recall (Tkn Recall) and accuracy (Tkn Acc%). Best performances are boldfaced.

Model	Disambiguation		Span Detection
Model	F1	Acc	Seq Acc	Tkn Recall	Tkn Acc
Majority Class	87.37	77.57	22.43	0.0	91.18
BART	95.89	93.71	50.76	75.45	96.51
BART-FT	96.46	94.49	61.53	84.98	97.24
ITI	96.04	93.88	55.07	79.16	96.82
ITI+SI	96.53	94.61	60.29	84.39	97.15
ITI+SF	95.81	93.52	54.97	76.75	96.69
ITI+SF+Copy	95.73	93.30	76.35	89.48	98.12
ITI+SF+SI	95.73	93.25	76.01	90.75	98.17

Model	Disambiguation		Span Detection
Model	F1	Acc	Seq Acc	Tkn Recall	Tkn Acc
Majority Class	87.37	77.57	22.43	0.0	91.18
BART	95.89	93.71	50.76	75.45	96.51
BART-FT	96.46	94.49	61.53	84.98	97.24
ITI	96.04	93.88	55.07	79.16	96.82
ITI+SI	96.53	94.61	60.29	84.39	97.15
ITI+SF	95.81	93.52	54.97	76.75	96.69
ITI+SF+Copy	95.73	93.30	76.35	89.48	98.12
ITI+SF+SI	95.73	93.25	76.01	90.75	98.17