An Improved Sign Language Translation Model with Explainable Adaptations for Processing Long Sign Sentences

Table 2

Experiments on the existing baseline systems vs. variants of our novel model.

#	Model	Development set					Test set
		ROUGE	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE	BLEU-1	BLEU-2	BLEU-3	BLEU-4
Existing baseline systems
2a	None	29.54	28.33	15.71	10.32	8.57	28.60	26.65	15.02	10.27	8.24
2b	Transformer	30.28	29.82	16.98	11.89	8.93	29.89	29.45	16.72	11.78	8.82
2c	Luong	31.67	32.18	18.56	12.38	9.46	30.71	30.01	17.43	12.11	9.02
2d	Bahdanau	31.93	31.66	18.70	12.79	9.53	31.56	31.32	18.36	12.36	9.25
Our proposed systems
2e	+T-Conv	32.08	30.08	18.15	12.88	9.97	31.34	30.94	18.26	12.71	9.76
2f	+DH-BiGRUs	31.55	30.21	18.29	13.05	9.84	31.20	31.46	17.64	12.40	9.65
2g	+TC-DHBG-Net (+T-Conv + DH-BiGRUs)	31.69	31.23	18.62	13.15	10.16	32.25	32.19	19.38	13.71	10.66
2h	+FSDC	32.13	31.72	18.84	12.98	9.79	31.52	31.72	19.04	13.01	9.71
2i	+FSDC + TC-DHBG-Net	32.76	31.43	19.12	13.40	10.35	32.99	31.86	19.51	13.81	10.73

Bold indicates the best performance.