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

Table 6

Performance of Transformer-based models on ParaNMT datasets.

Dataset	Model	BLEU	ROUGE-1	ROUGE-2	ROUGE-L
ParaNMT50K	base	13.7 (±0.47)	41.7 (±0.66)	19.29 (±0.5)	39.54 (±0.65)
	add	14.04 (±0.2)	42.37 (±0.15)^††	19.68 (±0.21)	40.18 (±0.17)^††
	cat	14.05 (±0.17)	42.55 (±0.23)^††	19.77 (±0.17)	40.34 (±0.24)^††
	dc	13.84 (±0.07)	41.97 (±0.09)	19.4 (±0.06)	39.78 (±0.08)
ParaNMT100K	base	15.8 (±0.23)	45.5 (±0.21)	22.01 (±0.21)	43.03 (±0.22)
	add	15.85 (±0.19)	45.76 (±0.3)^†	22.14 (±0.25)	43.3 (±0.27)^†
	cat	15.94 (±0.06)	46.09 (±0.17)^†††	22.34 (±0.09)^†††	43.58 (±0.12)^†††
	dc	15.38 (±0.1)	45.28 (±0.06)	21.69 (±0.09)	42.76 (±0.08)