A Comparative Study of Some Automatic Arabic Text Diacritization Systems

Table 5

Comparison of word error rate (WER) and diacritic error rate (DER).

Systems	Diacritic error rate: DER (%)				Word error rate: WER (%)
	Including ‘no diacritic’ class		Excluding ‘no diacritic’ class		Including ‘no diacritic’ class		Excluding ‘no diacritic’ class
	With case ending	Without case ending	With case ending	Without case ending	With case ending	Without case ending	With case ending	Without case ending
Mohamed Hadj Ameur et al.‘s work [21] (word-level HMM using Absolute discount smoothing technique)	10.29	7.78	12.09	9.12	23.37	11.44	22.49	11.12
Mohamed Hadj Ameur et al.‘s work [21] (word-level & letter-level HMMs using Absolute discount smoothing technique)	9.43	6.73	10.69	7.46	23.43	11.41	22.47	11.07
Mohamed Hadj Ameur et al.‘s work [21] (word-level HMM using Laplace smoothing technique)	10.27	7.94	12.02	9.28	22.76	11.53	21.90	11.24
Mohamed Hadj Ameur et al.‘s work [21] (word-level & letter-level HMMs using Laplace smoothing technique)	9.42	6.90	10.60	7.59	22.82	11.49	21.89	11.17
Hamza Abbad et al.‘s work [22] (using two LSTM layers)	11.52	10.70	13.02	11.91	29.62	22.58	27.61	21.02
Hamza Abbad et al.‘s work [22] (replacing the LSTM layers with two GRU layers)	12.53	11.41	14.21	12.71	31.43	23.50	29.53	22.07
Hamza Abbad et al.‘s work [22] (replacing only the second LSTM layer with a GRU layer)	12.10	11.09	13.69	12.34	30.66	22.98	28.81	21.60
Hamza Abbad et al.‘s work [22] (replacing only the first LSTM layer with a GRU layer)	11.81	10.94	13.37	12.18	30.14	22.79	28.16	21.26
Our proposed modified version of Hamza Abbad et al.‘s work (using LSTM layers)	12.11	10.97	13.73	12.20	31.46	23.15	29.45	21.59
Our proposed modified version of Hamza Abbad et al.‘s work (using GRU layers instead of LSTM layers)	13.21	11.62	15.03	12.97	33.57	24.21	31.53	22.65
Our proposed modified version of Hamza Abbad et al.‘s work (using an LSTM layer followed by a GRU layer)	12.63	11.31	14.36	12.61	32.55	23.81	30.56	22.27
Our proposed modified version of Hamza Abbad et al.‘s work (using a GRU layer followed by an LSTM layer)	12.50	11.21	14.21	12.51	32.18	23.61	30.20	22.08