A Comprehensive Survey of Abstractive Text Summarization Based on Deep Learning

Table 4

The results of different models on the Gigaword dataset. RG-1 denotes the ROUGE-1 score, RG-2 denotes ROUGE-2 score, and RG-L denotes ROUGE-L score.

Year	Method	Gigaword			Vocabulary
		RG-1	RG-2	RG-L	In/out
2017	SEASS [119]	36.15	17.54	33.63	120k/69k
	DRGD [120]	36.27	17.57	33.62	110k/69k
	FTSumg [100]	37.27	17.65	34.24	120k/69k
	Transformer [121]	37.57	18.90	34.69	120k/69k
2018	Struct + 2Way + Word [122]	35.47	17.66	33.52	70k/10k
	PG + EntailGen + QuestionGen [123]	35.98	17.76	33.63	110k/69k
	CGU [124]	36.3	18.0	33.8	110k/69k
	Reinforced-topic-ConvS2S [85]	36.92	18.29	34.58	110k/69k
	Seq2seq + E2T_cnn [125]	37.04	16.66	34.93	50k/50k
	Re^3 sum [126]	37.04	19.03	34.46	110k/69k
2019	JointParsing [127]	36.61	18.85	34.33	110k/69k
	Concept pointer + DS [128]	37.01	17.10	34.87	150k/150k
	MASS [129]	38.73	19.71	35.96	110k/69k
	UniLM [130]	38.90	20.05	36.00	30k/30k
	BiSET [131]	39.11	19.78	36.87	110k/69k
	PEGASUS [132]	39.12	19.86	36.24	96k/96k
2020	ERNIE-GENBASE [133]	38.83	20.04	36.20	50k/50k
	ERNIE-GENLARGE [133]	39.25	20.25	36.53	50k/50k
	ProphetNet [134]	39.51	20.42	36.69	110k/69k
	BART-RXF [135]	40.45	20.69	36.56	120k/69k
2021	Mask attention network [136]	38.28	19.46	35.46	110k/69k
	Transformer + Wdrop [137]	39.66	20.45	36.59	32k/32k
	Transformer + Rep [137]	39.81	20.40	36.93	32k/32k
	MUPPET BART large [138]	40.4	20.54	36.21	120k/69k

The values in bold represent the SOTA model for that year.