SentMask: A Sentence-Aware Mask Attention-Guided Two-Stage Text Summarization Component
Table 4
The human evaluation results. The score is calculated on an average of the scores for 300 news articles from the MS2 and AESLC datasets that were supplied by 5 volunteers. The score of each volunteer, which goes from 1 to 5, is the assessment of every news article.
Models
MS2 dataset
AESLC dataset
INFOR
FAITH
FLU
INFOR
FAITH
FLU
Semisupervised
TextRank
3.584
3.5666
3.5726
2.5893
2.572
2.582
GenCompareSum
3.6293
3.6306
3.2687
2.6493
2.6506
2.646
TextRank + Seq2Seq
3.6266
3.6393
3.636
2.6406
2.644
2.6526
SentMask
3.8326
3.8393
3.8433
2.8106
2.8233
2.8273
Supervised
AESLC
3.682
3.6853
3.6833
2.7033
2.7066
2.7046
Transformer
3.816
3.862
3.8366
2.8533
2.8166
2.8206
Pointer-gen
3.8146
3.83
3.822
2.8066
2.862
2.8586
BART
3.8693
3.8833
3.8533
2.88
2.8873
2.8746
SentMask
3.9333
3.9373
3.9406
2.9466
2.954
2.9493
The best values in the metric are in bold. INFOR: informativeness; FAITH: faithfulness; FLU: fluency.
