MIT Press

Table 9:

Ranking parsers: STOG (Zhang et al.); GPLA (Lyu and Titov); TOP-DOWN (Cai and Lam, 2019). The structure error is defined as

\frac{1}{1371} Σ_{i = 1}^{1371} | f (g o l d_{i}) - f (p r e d_{i}) |

⁠, where f either is node degree or graph density. All four metrics differ significantly in their scores (paired t-test, p<0.05).

	metric scores				structure error
	Sm	${SB}_{A}^{G}$	${SB}_{G}^{A}$	S²m	node degree	graph density
STOG	76.3_\|1	59.6_\|1	58.9_\|1	77.9_\|1	0.08	0.0069
GPLA	74.5_\|2	54.2_\|3	52.9_\|3	76.2_\|2	0.08	0.0068
TOP-DOWN	73.2_\|3	54.5_\|2	53.1_\|2	75.0_\|3	0.11	0.0078

	metric scores				structure error
	Sm	${SB}_{A}^{G}$	${SB}_{G}^{A}$	S²m	node degree	graph density
STOG	76.3_\|1	59.6_\|1	58.9_\|1	77.9_\|1	0.08	0.0069
GPLA	74.5_\|2	54.2_\|3	52.9_\|3	76.2_\|2	0.08	0.0068
TOP-DOWN	73.2_\|3	54.5_\|2	53.1_\|2	75.0_\|3	0.11	0.0078