MIT Press

Table 3

Results for models of different context representations: Char(acter), W(ord)P(iece), W(ord)F(eature). For the contextual models we specify the tokenization level of the sentential context. *: difference is statistically significant (p < .05 on the McNemar test) in comparison with the sliding window baseline. For speed, e.g., “1.4x” means 1.4 times faster.

	Accuracy (Error)			Speed
	All	Semiotic class	Sentence	Speed
English (Standard):
Sliding window	99.79% (0.21%)	98.20% (1.80%)	97.99% (2.01%)	1.0x
Context (Char)	99.79% (0.21%)	98.20% (1.80%)	97.87% (2.13%)	1.3x
Context (WP)	99.79% (0.21%)	98.35% (1.65%)	97.76% (2.24%)	1.4x
Context (WF)	99.84% (0.16%)*	98.30% (1.70%)	98.20% (1.80%)	1.4x

Russian (Standard):
Sliding window	99.64% (0.36%)	97.31% (2.69%)	95.61% (4.39%)	1.0x
Context (Char)	99.65% (0.35%)	97.26% (2.74%)	95.70% (4.30%)	1.8x
Context (WP)	99.61% (0.39%)	96.94% (3.06%)*	95.26% (4.74%)	1.8x
Context (WF)	99.62% (0.38%)	97.01% (2.99%)*	95.46% (4.54%)	2.0x

	Accuracy (Error)			Speed
	All	Semiotic class	Sentence	Speed
English (Standard):
Sliding window	99.79% (0.21%)	98.20% (1.80%)	97.99% (2.01%)	1.0x
Context (Char)	99.79% (0.21%)	98.20% (1.80%)	97.87% (2.13%)	1.3x
Context (WP)	99.79% (0.21%)	98.35% (1.65%)	97.76% (2.24%)	1.4x
Context (WF)	99.84% (0.16%)*	98.30% (1.70%)	98.20% (1.80%)	1.4x

Russian (Standard):
Sliding window	99.64% (0.36%)	97.31% (2.69%)	95.61% (4.39%)	1.0x
Context (Char)	99.65% (0.35%)	97.26% (2.74%)	95.70% (4.30%)	1.8x
Context (WP)	99.61% (0.39%)	96.94% (3.06%)*	95.26% (4.74%)	1.8x
Context (WF)	99.62% (0.38%)	97.01% (2.99%)*	95.46% (4.54%)	2.0x

View Large

This Feature Is Available To Subscribers Only