Hyper-parameter optimization in neural-based translation systems: A case study

Datta, Goutam; Joshi, Nisheeth; Gupta, Kusum

doi:10.2478/ijssis-2023-0010

Figures & Tables

Schematic illustration of computation of ∂j for hidden unit j with the help of back propagation.

The accuracy of deep learning models is higher, but their interpretability is low compared to other ML models.

Schematic representation of bidirectional RNN (Source: Bahdanau et al. [26]).

The generative adversarial network (GAN) model in the NMT case.

Typical machine learning model building steps.

Schematic representation of the NMT model.

Training and validation accuracy with five epochs.

There is a slight improvement in training and validation accuracy with increasing the number of epochs up to 10.

Graphical representation of training and validation accuracy with reduced units in different layers and up to five epochs.

Graphical representation of training and validation accuracy with reduced units per layer and increased number of epochs, i.e., up to 10.

Snapshots of the English–Bangla parallel corpus collected from TDIL.

BLEU scores produced by different NMT models for the first test data.

BLEU scores generated by different NMT models for the second test data.

BLEU score produced by various NMT models on third test data.

WMT-14 English–German test results show that ADMIN outperforms the default base model 6L-6L in different automatic metrics (Liu et al_ [35])_

Model	Param	TER	METEOR	BLEU
6L-6L Default	61M	54.4	46.6	27.6
6L-6L ADMIN	61M	54.1	46.7	27.7
60L-12LDefault	256M	Diverge	Diverge	Diverge
60L-12LADMIN	256M	51.8	48.3	30.1

Statistics of English to Bangla tourism corpus (text) collected from TDIL_

Corpus (English to Bangla)	Size in terms of sentence pairs
Tourism	11,976

NMT models with some other range of learning rate (hyper-parameter) (Lim et al_ [11])_

Cell	Learning rate	ro→en P100	ro→en t	ro→en V100	ro→en t	de→en P100	de→en t	de→en V100	de→en t
GRU	0.0	34.47	6:29	34.47	4:43	32.29	9:48	31.61	6:15
	0.2	35.53	8:48	35.43	6:21	33.03	18:47	32.55	19:40
	0.3	35.36	12:21	35.15	7:28	31.36	10:14	31.50	9:33
	0.5	34.50	12:20	34.67	17:18	29.64	11:09	30.21	11.09
LSTM	0.0	34.84	6:29	34.65	4:46	32.84	12:17	32.88	7:37
	0.2	34.27	8:10	35.61	6:34	33.10	16:33	33.89	13:39
	0.3	35.67	9:56	35.37	11:29	33.45	20.02	33.51	15:51
	0.5	34.50	15:13	34.33	12:45	32.67	20.02	32.20	13.03

Training and validation accuracy of our model with five epochs_

Epochs	Training accuracy	Validation accuracy
1	0.9426	0.9698
2	0.9730	0.9708
3	0.9792	0.9776
4	0.9829	0.9726
5	0.9859	0.9762

Translations generated by Google and Bing_

Translators	Language pair	BLEU
Google	English ⇒ Bangla (1^st sentence)	36.84
	English ⇒ Bangla (2^nd sentence)	6.42
	English ⇒ Bangla (3^rd sentence)	4.52
Bing	English ⇒ Bangla (1^st sentence)	36.11
	English ⇒ Bangla (2^nd sentence)	6.01
	English ⇒ Bangla (3^rd sentence)	4.05

Training and validation accuracy with 100 units in different layers with 10 epochs_

Epochs	Training accuracy	Validation accuracy
1	0.9293	0.9631
2	0.9674	0.9730
3	0.9763	0.9751
4	0.9807	0.9729
5	0.9829	0.9724
6	0.9852	0.9780
7	0.9882	0.9773
8	0.9890	0.9756
9	0.9908	0.9784
10	0.9913	0.9793

Generic hyper-parameters in NMT-based model_

Model	Type of MT	Hyper-parameters
Deep learning models	NMT	Hidden layers, learning rate, activation function, epochs, batch size, dropout, regularization

Models per data set and their best BLEU scores and respective hyper-parameter configurations (Zhang and Duh [36])_

Data set	No. of models	Best BLEU	BPE	No. of layers	No. of embedding	No. of hidden layers	No. of attention heads	Init-lr
Chinese–English	118	14.66	30k	4	512	1024	16	3e-4
Russian–English	176	20.23	10k	4	256	2048	8	3e-4
Japanese–English	150	16.41	30k	4	512	2048	8	3e-4
English–Japanese	168	20.74	10k	4	1024	2048	8	3e-4
Swahili–English	767	26.09	1k	2	256	1024	8	6e-4
Somali–English	604	11.23	8k	2	512	1024	8	3e-4

Training and validation accuracy of our model with a higher number of epochs_

Epochs	Training accuracy	Validation accuracy
1	0.9431	0.9606
2	0.9742	0.9729
3	0.9796	0.9777
4	0.9835	0.9748
5	0.9865	0.9794
6	0.9872	0.9802
7	0.9896	0.9830
8	0.9898	0.9782
9	0.9916	0.9764
10	0.9924	0.9799

WMT-14 English–French test results showed that 60L-12L ADMIN outperforms the default base model 6L-6L in different automatic metrics (Liu et al_ [35])_

Model	Param	TER	METEOR	BLEU
6L-6L Default	67M	42.2	60.5	41.3
6L-6L ADMIN	67M	41.8	60.7	41.5
60L-12LDefault	262M	Diverge	Diverge	Diverge
60L-12LADMIN	262M	40.3	62.4	43.8

MT models for different language pairs in a GPU-based single-node and multiple-node environment with a wider range of hyper-parameters and their BLEU scores (Lim et al_ [11])_

Cell	Learning rate	ro→en P100	ro→en V100	en→ro P100	en→ro V100	de→en P100	de→en V100	en→de P100	en→de V100
GRU	le-3	35.53	35.43	19.19	19.28	28.00	27.84	20.43	20.61
	5e-3	34.37	34.05	19.07	19.16	26.05	22.16	N/A	19.01
	le-4	35.47	35.46	19.45	19.49	27.37	27.81	Dnf	21.41
LSTM	le-3	34.27	35.61	19.29	19.64	28.62	28.83	21.70	21.69
	5e-3	35.05	34.99	19.48	19.43	N/A	24.36	18.53	18.01
	le-4	35.41	35.28	19.43	19.48	N/A	28.50	Dnf	Dnf
GRU	le-3	34.22	34.17	19.42	19.43	33.03	32.55	26.55	26.85
	5e-3	33.13	32.74	19.31	18.97	31.04	26.76	N/A	26.02
	le-4	33.67	34.44	18.98	19.69	33.15	33.12	Dnf	28.43
LSTM	le-3	33.10	33.95	19.56	19.08	33.10	33.89	28.79	28.84
	5e-3	33.10	33.52	19.13	19.51	N/A	29.16	24.12	24.12
	le-4	33.29	32.92	19.14	19.23	N/A	33.44	Dnf	Dnf

Training and validation accuracy with 100 units in different layers with five epochs_

Epochs	Training accuracy	Validation accuracy
1	0.9289	0.9584
2	0.9674	0.9671
3	0.9758	0.9734
4	0.9800	0.9739
5	0.9836	0.9772

Performance of BiLSTM, Google Translate, and Bing in terms of the automatic metric BLEU_

Model	Hyper-parameter	BLEU score
BiLSTM (for English to Bangla; 1^st sentence)	Optimizer = Adam;	4.1
BiLSTM (for English to Bangla; 2^nd sentence)	Learning rate = 0.001;	3.2
BiLSTM (for English to Bangla; 3^rd sentence)	No. of encoder and decoder layers = 6	3.01

Hyper-parameter optimization in neural-based translation systems: A case study

Figures & Tables

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Figure 6:

Figure 7:

Figure 8:

Figure 9:

Figure 10:

Figure 11:

Figure 12:

Figure 13:

Figure 14:

Figure 15:

Figure 16:

Figure 17:

Figure 18:

WMT-14 English–German test results show that ADMIN outperforms the default base model 6L-6L in different automatic metrics (Liu et al_ [35])_

Statistics of English to Bangla tourism corpus (text) collected from TDIL_

NMT models with some other range of learning rate (hyper-parameter) (Lim et al_ [11])_

Training and validation accuracy of our model with five epochs_

Translations generated by Google and Bing_

Training and validation accuracy with 100 units in different layers with 10 epochs_

Generic hyper-parameters in NMT-based model_

Models per data set and their best BLEU scores and respective hyper-parameter configurations (Zhang and Duh [36])_

Training and validation accuracy of our model with a higher number of epochs_

WMT-14 English–French test results showed that 60L-12L ADMIN outperforms the default base model 6L-6L in different automatic metrics (Liu et al_ [35])_

MT models for different language pairs in a GPU-based single-node and multiple-node environment with a wider range of hyper-parameters and their BLEU scores (Lim et al_ [11])_

Training and validation accuracy with 100 units in different layers with five epochs_

Performance of BiLSTM, Google Translate, and Bing in terms of the automatic metric BLEU_

Paradigm

My account