DEplain: Automatic Text Simplification Experiments

July 1, 2023 · View on GitHub

To exemplify the usage of DEPLAIN for training and evaluating TS models, we are presenting results on finetuning long-mBART on our document-level corpus as well as finetuning mBART on our sentence-level corpus, using code provided by Rios et al. (2021).

You can find our edited code here: https://github.com/omarTronto/ats-models.

Document Simplification

You can find the generated simplifications of DEplain-web here: . The generated files of DEplain-APA are stored on zenodo, please request access to it.

Finetuned models checkpoints

Results on Document-level

train data	n	SARI ↑	BLEU ↑	BS-P ↑	FRE ↑
DEplain-APA	387	44.56	38.136	0.598	65.4
DEplain-web	481	35.02	12.913	0.475	59.55
DEplain-APA+web	868	42.862	36.449	0.589	65.4
src2Src-baseline		17.637	34.247	0.583	58.85

: Results on DEplain-APA test (n=48).

train data	n	SARI ↑	BLEU ↑	BS-P ↑	FRE ↑
DEplain-APA	387	43.087	21.9	0.377	64.7
DEplain-web	481	49.584	23.282	0.462	63.5
DEplain-APA+web	868	49.745	23.37	0.445	57.95
src2Src-baseline		12.848	23.132	0.432	59.4

: Results on DEplain-web test (n=147).

Sentence Simplification

You can find the generated simplifications of DEplain-web here: . The generated files of DEplain-APA are stored on zenodo, please request access to it.

Finetuned models checkpoints

Results on Sentence-level

train data	n	SARI ↑	BLEU ↑	BS-P ↑	FRE ↑
DEplain-APA	10660	34.818	28.25	0.639	63.072
DEplain-APA+web	11941	34.904	28.506	0.64	62.669
src2src-baseline		15.249	26.893	0.627	59.23

: Results on DEplain-APA test (n=1231).

train data	n	SARI ↑	BLEU ↑	BS-P ↑	FRE ↑
DEplain-APA	10660	30.867	15.727	0.413	64.516
DEplain-APA+web	11941	34.828	17.88	0.436	65.249
src2src-baseline		11.931	20.85	0.423	60.825

: Results on DEplain-web test (n=1846).