MIT Press

Table 2:

Performance comparison among traditional, state-of-the-art, ablation strategies, and Vec2Read on different data sets. ‘*’ denotes statistically significant improvement over counterparts (Flesch, S1, S2, S3, Vec2Read). Accuracy (higher is better) is reported for all data sets except for MTDE, where RMSE (lower is better) is used in order to be able to compare with S1. Cells marked with ‘-’ denote that the strategy is not applicable to the data set.

Data set	Lang.	Flesch	State of the art			Ablation							Vec2Read
			S1	S2	S3	FC	¬Attention	¬Word	¬Sent	¬Sem	¬Syn	¬Morph
Binary Prediction (Accuracy)
SimpleWiki	en	.724	-	-	.822	.722	.877	.893	.896	.887	.897	.915	.918^⋆
VikiWiki	en	.720	-	-	.827	.721	.852	.860	.862	.859	.868	.876	.879^⋆
	es	.687	-	-	.792	.719	.816	.823	.831	.828	.835	.839	.847^⋆
	fr	.670	-	-	.842	.756	.864	.869	.870	.869	.870	.872	.884^⋆
	it	.653	-	-	.755	.766	.783	.797	.802	.793	.801	.805	.814^⋆
	eu	-	-	-	.693	.648	.682	.683	.686	.684	.684	.685	.687^⋆
	ca	-	-	-	.733	.677	.715	.725	.737	.728	.732	.734	.742^⋆

Multilevel Prediction (Accuracy)
Ikasbil	eu	-	-	.625	.622	.617	.679	.685	.689	.681	.684	.686	.692^⋆
Newsela	en	-	-	-	.464	.447	.489	.501	.517	.498	.502	.525	.527^⋆
	es	-	-	-	.467	.452	.487	.494	.510	.504	.509	.503	.519^⋆
Wizenoze	en	-	-	-	.649	.631	.665	.678	.685	.682	.685	.700	.701^⋆
	du	-	-	-	.652	.636	.668	.679	.687	.681	6.85	.683	.695^⋆

Continuous Prediction (RMSE)
MTDE	du	-	.0003^⋆	-	-	.0171	.0068	.0064	.0064	.0066	.0062	.0059	.0059^⋆
	en	-	.0060^⋆	-	-	.0184	.0054	.0052	.0051	.0051	.0053	.0051	.0051^⋆

Data set	Lang.	Flesch	State of the art			Ablation							Vec2Read
			S1	S2	S3	FC	¬Attention	¬Word	¬Sent	¬Sem	¬Syn	¬Morph
Binary Prediction (Accuracy)
SimpleWiki	en	.724	-	-	.822	.722	.877	.893	.896	.887	.897	.915	.918^⋆
VikiWiki	en	.720	-	-	.827	.721	.852	.860	.862	.859	.868	.876	.879^⋆
	es	.687	-	-	.792	.719	.816	.823	.831	.828	.835	.839	.847^⋆
	fr	.670	-	-	.842	.756	.864	.869	.870	.869	.870	.872	.884^⋆
	it	.653	-	-	.755	.766	.783	.797	.802	.793	.801	.805	.814^⋆
	eu	-	-	-	.693	.648	.682	.683	.686	.684	.684	.685	.687^⋆
	ca	-	-	-	.733	.677	.715	.725	.737	.728	.732	.734	.742^⋆

Multilevel Prediction (Accuracy)
Ikasbil	eu	-	-	.625	.622	.617	.679	.685	.689	.681	.684	.686	.692^⋆
Newsela	en	-	-	-	.464	.447	.489	.501	.517	.498	.502	.525	.527^⋆
	es	-	-	-	.467	.452	.487	.494	.510	.504	.509	.503	.519^⋆
Wizenoze	en	-	-	-	.649	.631	.665	.678	.685	.682	.685	.700	.701^⋆
	du	-	-	-	.652	.636	.668	.679	.687	.681	6.85	.683	.695^⋆

Continuous Prediction (RMSE)
MTDE	du	-	.0003^⋆	-	-	.0171	.0068	.0064	.0064	.0066	.0062	.0059	.0059^⋆
	en	-	.0060^⋆	-	-	.0184	.0054	.0052	.0051	.0051	.0053	.0051	.0051^⋆

View Large

This Feature Is Available To Subscribers Only