MIT Press

Table 3:

Development results of the variations on Graph Transformer. Unlisted values are identical to those of the base model. Both models trained on LDC2015E86 and LDC2017T10 are evaluated.

	L₁	L₂	d_model	d_h	P_drop	BLEU	BLEU
						(LDC2015E86)	(LDC2017T10)
base	8	6	256	2	0.3	25.5	28.8

(A)	2					20.4	24.6
	4					23.7	27.6
	6					24.3	28.3
	10					24.6	28.4

(B)		4				23.4	27.1
(B)		5				24.6	28.0
		7				24.8	28.7

(C)			512			25.1	28.5

(D)				1		23.6	28.4
(D)				4		23.0	28.7

(E)					0.1	22.7	26.9
(E)					0.2	25.3	28.5
					0.4	24.7	27.9

(F)	single representation for each node					25.1	28.3
(F)	single representation, inseparate graph attention					23.1	27.6

	L₁	L₂	d_model	d_h	P_drop	BLEU	BLEU
						(LDC2015E86)	(LDC2017T10)
base	8	6	256	2	0.3	25.5	28.8

(A)	2					20.4	24.6
	4					23.7	27.6
	6					24.3	28.3
	10					24.6	28.4

(B)		4				23.4	27.1
(B)		5				24.6	28.0
		7				24.8	28.7

(C)			512			25.1	28.5

(D)				1		23.6	28.4
(D)				4		23.0	28.7

(E)					0.1	22.7	26.9
(E)					0.2	25.3	28.5
					0.4	24.7	27.9

(F)	single representation for each node					25.1	28.3
(F)	single representation, inseparate graph attention					23.1	27.6

View Large

This Feature Is Available To Subscribers Only