Dou et al. (2021)#
Publication#
GSum: A General Framework for Guided Neural Abstractive Summarization
Repositories#
https://github.com/neulab/guided_summarization
Available Models#
The authors have released their BART-based model that uses sentence guidance, downloadable here.
Oracle Sentence-Guided
Description: A BART-based model trained with sentence supervision. This model will use the oracle supervision, which is calculated by selecting document sentences which maximize the ROUGE score with respect to the ground-truth. Therefore, it requires the reference summaries as input.
Name:
dou2021-oracle-sentence-gsum
Usage:
from repro.models.dou2021 import OracleSentenceGSumModel model = OracleSentenceGSumModel() summary = model.predict("document", "reference")
Sentence-Guided
Description: A BART-based model trained with sentence supervision. This is the same pre-trained model as the oracle sentence-guided supervision, but it instead uses
BertSumExt
to extract the sentence guidance. Running this model also requires having theBertSumExt
model setup with the model pre-trained on CNN/DailyMail. We caution that theBertSumExt
currently does not reproduce the results in the respective paper.Name:
dou2021-sentence-gsum
Usage:
from repro.models.dou2021 import SentenceGSumModel model = SentenceGSumModel() summary = model.predict("document")
Implementation Notes#
Docker Information#
Image name:
dou2021
Build command:
repro setup dou2021 [--silent]
Requires network: Yes.
fairseq
sends a request to retrieve an etag even if the file is present. This also happens with Lewis et al. (2020).
Testing#
repro setup dou2021
pytest models/dou2021/tests
Status#
[x] Regression unit tests pass
Only the tests which require models from Liu & Lapata (2019) fail on Github. See here.[ ] Correctness unit tests pass
[x] Model runs on full test dataset
See here[x] Predictions approximately replicate results reported in the paper
The MatchSum-guided BART-based model was the only BART-based result reported in the paper, and it does come sufficiently close.R1
R2
RL
Reported
45.94
22.32
42.48
Ours
45.80
22.18
42.44
See here for more details about the different model variants.
[ ] Predictions exactly replicate results reported in the paper
Misc#
See these notes if this code is extended to include training.