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

Table 4

Performance of RNN-based models on COCO datasets.

Dataset	Model	BLEU	ROUGE-1	ROUGE-2	ROUGE-L
COCO20K	base	6.15 (±0.23)	35.32 (±0.22)	10.86 (±0.23)	29.71 (±0.18)
	add	6.2 (±0.14)	35.77 (±0.11)^†††	11.01 (±0.06)	30.03 (±0.07)^†††
	cat	6.27 (±0.21)	35.74 (±0.15)^†††	11.04 (±0.18)	30.01 (±0.17)^††
	dc	6.4 (±0.05)^††	35.8 (±0.26)^††	11.46 (±0.15)^†††	30.36 (±0.27)^†††
COCO50K	base	6.9 (±0.04)	37.41 (±0.2)	12.02 (±0.16)	31.42 (±0.12)
	add	6.99 (±0.2)	37.54 (±0.14)	12.19 (±0.07)^††	31.6 (±0.09)^††
	cat	6.92 (±0.19)	37.6 (±0.09)^†	12.11 (±0.12)	31.58 (±0.08)^††
	dc	6.99 (±0.1)^†	37.56 (±0.21)	12.44 (±0.18)^†††	31.76 (±0.19)^†††
COCO	base	7.68 (±0.03)	39.18 (±0.17)	13.11 (±0.1)	32.95 (±0.09)
	add	7.61 (±0.14)	39.15 (±0.18)	13.13 (±0.11)	32.91 (±0.12)
	cat	7.55 (±0.15)	39.13 (±0.15)	13.02 (±0.13)	32.89 (±0.14)
	dc	7.62 (±0.09)	39.21 (±0.15)	13.3 (±0.13)^††	33.03 (±0.05)^†