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

Table 5

Performance of Transformer-based models on Quora datasets.

Dataset	Model	BLEU	ROUGE-1	ROUGE-2	ROUGE-L
Quora10K	base	20.11 (±0.11)	48.84 (±0.17)	27.28 (±0.17)	47.57 (±0.18)
	add	20.73 (±0.15)	49.82 (±0.16)	28.17 (±0.16)	48.56 (±0.16)
	cat	20.92 (±0.11)	50.47 (±0.26)	28.72 (±0.22)	49.17 (±0.25)
	dc	20.98 (±0.09)	49.49 (±0.15)	28.12 (±0.16)	48.26 (±0.15)
Quora20K	base	22.07 (±0.12)	52.09 (±0.21)	29.91 (±0.2)	50.53 (±0.18)
	add	22.41 (±0.12)	52.82 (±0.13)	30.46 (±0.11)	51.21 (±0.14)
	cat	22.71 (±0.23)	53.24 (±0.21)	30.87 (±0.23)	51.65 (±0.2)
	dc	22.64 (±0.18)	52.43 (±0.2)^††	30.39 (±0.2)	50.92 (±0.21)^††
Quora50K	base	24.3 (±0.1)	56.02 (±0.08)	33.07 (±0.09)	54.1 (±0.07)
	add	24.84 (±0.14)	56.58 (±0.18)	33.59 (±0.16)	54.66 (±0.18)
	cat	24.95 (±0.26)	56.8 (±0.07)	33.86 (±0.1)	54.88 (±0.04)
	dc	24.71 (±0.15)	56.25 (±0.11)	33.44 (±0.08)	54.41 (±0.09)