Fusing Part-of-Speech Information in Low-Resource Neural Paraphrase Generation

Table 7

Performance of Transformer-based models on COCO datasets.

Dataset	Model	BLEU	ROUGE-1	ROUGE-2	ROUGE-L
COCO20K	base	7.85 (±0.1)	38.18 (±0.09)	13.94 (±0.07)	35.33 (±0.05)
	add	8.05 (±0.11)^††	38.4 (±0.18)^††	14.16 (±0.13)^†††	35.49 (±0.15)^††
	cat	8.16 (±0.1)^†††	38.65 (±0.17)^†††	14.29 (±0.06)^†††	35.57 (±0.2)^††
	dc	7.88 (±0.11)	38.27 (±0.08)^†	13.96 (±0.11)	35.28 (±0.11)
COCO50K	base	8.34 (±0.09)	39.79 (±0.08)	14.84 (±0.05)	36.48 (±0.04)
	add	8.5 (±0.16)^†	40.04 (±0.17)^††	15.07 (±0.12)^†††	36.68 (±0.11)^†††
	cat	8.56 (±0.08)^†††	40.0 (±0.06)^†††	15.08 (±0.06)^†††	36.58 (±0.06)^††
	dc	8.21 (±0.2)	39.69 (±0.21)	14.77 (±0.17)	36.43 (±0.11)
COCO	base	8.92 (±0.16)	41.15 (±0.28)	15.74 (±0.2)	37.63 (±0.19)
	add	9.1 (±0.09)^††	41.51 (±0.09)^††	15.98 (±0.1)^†††	37.83 (±0.08)^††
	cat	9.12 (±0.19)^†	41.46 (±0.24)^†	15.97 (±0.19)^†	37.8 (±0.17)
	dc	8.84 (±0.19)	41.13 (±0.26)	15.67 (±0.2)	37.58 (±0.2)