Performance Comparison of Statistical vs. Neural-Based Translation System on Low-Resource Languages

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Figure 6:

Figure 7:

Figure 8:

Figure 9:

Figure 10:

Attention-based NMT outperforms SMT for the Bengali–Hindi language pair (Das et al_ [32])

Translation model	BLEU score	Iterations
Attention-based NMT model	20.41	25
MOSES (SMT)	14.35	-

NMT outperformed SMT with transfer learning, ensemble, and further processing of data (Zopth et al_)

Language	SBMT	NMT	Transfer	Final
Hausa	23.7	16.8	21.3	24.0
Turkish	20.4	11.4	17.0	18.7
Uzbek	17.9	10.7	14.4	16.8
Urdu	17.9	5.2	13.8	14.5

NMT system with transformer model and BPE outperformed phrase-based SMT for English–Hindi and Hindi–English language pairs (Haque et al_ [33])

MT model	BLEU	METEOR	TER
Eng.Hindi-PBSMT	28.8	30.2	53.4
Eng.Hindi-NMT	36.6	33.5	46.3
Hindi–Eng.PBSMT	34.1	36.6	50.0
Hindi–Eng.NMT	39.9	38.5	42.0

English–Hindi translation using different optimizers

Language pair	Optimizer	BLEU-4 score	NMT model	No. of epochs
Eng.–Hindi	Adam	12.25	NMT with attention	14
Eng.–Hindi	SGD	11.50	NMT with attention	14
Eng.–Hindi		16.64	MOSES

English–Bengali translation BLEU scores using different optimizers

Language pairs	Optimizer	BLEU-4 score	MT model	No. of epochs
Eng.–Beng.	Adam	10.78	NMT with attention	12
Eng.–Beng.	SGD	11.17	NMT with attention	12
Eng.–Beng.		14.58	MOSES

BLEU-1, 2, and 3 scores are summarized for Eng_–Beng_ and Eng_–Hindi language pairs using Adam- and SGD-Optimizers

BLEU	Eng.–Beng.-NMT (Adam-Optimizer)	Eng.–Beng.-NMT (SGD-Optimizer)	Eng.–Hindi (NMT-Adam)	Eng.–Hindi (NMT-SGD)
BLEU-1	14.15	13.91	15.77	14.18
BLEU-2	12.65	13.11	14.12	13.33
BLEU-3	11.83	12.17	13.95	12.19

For various low-resource corpus SMT outperformed NMT (Ahmadnia et al_ [17])

Corpus	SMT	NMT	NMT*	NMT**
Gnome	20.54	15.49	17.26	18.76
KDE4	15.64	13.36	14.29	15.71
Subtitles	18.82	18.62	19.51	22.54
Ubuntu	16.76	14.27	15.14	15.87
Tanzil	17.69	15.14	16.53	17.72
Overall	17.06	15.25	16.67	17.32