We provide the code for the papers:
- "Entity-level Factual Consistency of Abstractive Text Summarization", EACL 2021.
- We provide a set of new metrics to quantify the entity-level factual consistency of generated summaries. We also provide code for the two methods in our paper:
- JAENS: joint entity and summary generation, and
- Summary-worthy entity classification with summarization (multi-task learning)
- We provide a set of new metrics to quantify the entity-level factual consistency of generated summaries. We also provide code for the two methods in our paper:
- "Improving Factual Consistency of Abstractive Summarization via Question Answering", ACL-IJCNLP 2021
- QUALS, a new automatic metric for factual consistency.
- CONSEQ, a new contrastive learning algorithm for Seq2seq models to optimize sequence level objectives such as QUALS.
Our code is based on the fairseq library and we added support for model training on Sagemaker.
python==3.6
:conda create -n entity_fact python=3.6
pytorch==1.4.0
:pip install torch==1.4.0 torchvision==0.5.0
- run
pip install --editable ./
- install
file2rouge
following instructions here - download
en_core_web_lg
:python -m spacy download en_core_web_lg
We provide three options to preprocess summarization data through the filter_level
option.
filter_level=0
: no special processingfilter_level=1
: remove corruption text in source articles and summaries. (Undesirable texts included as a result of imperfect data collection. e.g. "Share this with Email, Facebook, Messenger". Undesirable summaries such as "Collection of all USATODAY.com coverage of People, including articles, videos, photos, and quotes.")filter_level=2
: entity hallucination filtering in addition to corruption text removal. A summary sentence is removed if it contains a named entity not in the source document.
- Follow the instructions here to download and extract text from HTML files and establish the
xsum-extracts-from-downloads
directory. - Let
<xsum-data-dir>
be the directory that contains thexsum-extracts-from-downloads
directory andXSum-TRAINING-DEV-TEST-SPLIT-90-5-5.json
. - Run
python preprocess/data_prepro_clean.py --mode preprocess_xsum --input_dir <xsum-data-dir> --output_dir <xsum-data-dir>/processed-data --filter_level 0
withfilter_level
set to 0, 1, or 2.
- Download and unzip the stories directories from here for both CNN and Daily Mail. Put all .story files in a directory
<cnndm-data-dir>/raw_stories
. - Download the url files
mapping_train.txt
,mapping_test.txt
andmapping_valid.txt
from here to<cnndm-data-dir>
. - Run
python preprocess/data_prepro_clean.py --mode preprocess_cnndm --input_dir <cnndm-data-dir> --output_dir <cnndm-data-dir>/processed-data --filter_level 0
withfilter_level
set to 0, 1, or 2.
- Download the datasets following instructions from here.
- Run
python preprocess/data_prepro_clean.py --mode preprocess_newsroom --input_dir <newsroom-data-dir> --output_dir <newsroom-data-dir>/processed-data --filter_level 0
withfilter_level
set to 0, 1, or 2.
Download bpe encoder.json, vocabulary and fairseq dictionary to a directory, say <bpe-dir>
; then tokenize and binarize the data.
wget -O <bpe-dir>/encoder.json 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json'
wget -O <bpe-dir>/vocab.bpe 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe'
wget -N <bpe-dir>/dict.txt' https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt'
cd preprocess
python data_prepro_clean.py --mode bpe_binarize --input_dir <processed-data-dir> --tokenizer_dir <bpe-dir>
This generates the binary input files as well as the dictionaries under <processed-data-dir>/data_bin
for fairseq.
The idea is to train the seq2seq model to generate <summary-worthy named entities> <sep> <abstractive summary>
.
python preprocess/create_entity_classification_labels.py --base_dir <processed-data-dir> --type entity_augment --tokenizer_dir <bpe-dir>
Binarize the augmented targets:
cd preprocess
python data_prepro_clean.py --mode binarize --input_dir <processed-data-dir>/entity_augment --tokenizer_dir <bpe-dir>
Since we already binarized the source documents, we just need to create symbolic links to put all binary input files together for fairseq training:
ln -s <processed-data-dir>/data_bin/train.source-target.source.idx <processed-data-dir>/entity_augment/data_bin/train.source-target.source.idx
ln -s <processed-data-dir>/data_bin/train.source-target.source.bin <processed-data-dir>/entity_augment/data_bin/train.source-target.source.bin
ln -s <processed-data-dir>/data_bin/valid.source-target.source.bin <processed-data-dir>/entity_augment/data_bin/valid.source-target.source.bin
ln -s <processed-data-dir>/data_bin/valid.source-target.source.idx <processed-data-dir>/entity_augment/data_bin/valid.source-target.source.idx
Run the launch scripts scripts/launch_xsum.py
, scripts/launch_cnndm.py
or scripts/launch_newsroom.py
to fine-tune the BART-large model.
Note you need to modify the following in the scripts:
hyperparameters
.train_path
: location of the binary input files. e.g.<processed-data-dir>/entity_augment/data_bin
.init_path
: location of the pre-trained BART-large model checkpoint. Please rename the checkpoint topretrained_model.pt
output_path
: location for the model outputs.
If training locally, you need to specify ngpus
- the number of GPUS in the local machine. Example command:
python scripts/launch_xsum.py --datatype ner_filtered --epoch 8 --exp_type local
If training on Sagemaker, you need to specify the docker image name (image_name
) as well as execution role (role
).
To create Sagemaker docker container and push to ECR:
./build_and_push.sh <YOUR ECR REPO>
To launch training job:
python scripts/launch_xsum.py --datatype ner_filtered --epoch 8 --exp_type sagemaker
preprocess/multi_gpu_generate.py
is used to generate summaries.
Since the JAENS models generates the named entities before the summaries, we need to remove the named entities before evaluating the summaries. Example command:
python evaluate_hypo.py --mode remove_ent_from_hypo --base_dir <output-dir> --sub_dir <output-sub-dir> --split val --pattern .*.hypo
We use the tokenizer from Stanford CoreNLP package. Example command:
export CLASSPATH=path/to/stanford-corenlp-full-2018-10-05/stanford-corenlp-3.9.2.jar
python evaluate_hypo.py --mode evaluate_summary --base_dir <output-dir> --sub_dir <output-sub-dir> --split val --pattern .*.hypo
See preprocess/run_*_eval.sh
for examples.
We perform summary-worthy entity classification at a classification head on the encoder while keeping the seq2seq objective at the decoder. For training, we need to preprocess the input document by create B-I-O labels to identify summary-worthy entities:
python create_entity_classification_labels.py --base_dir <processed-data-dir> --type cls_labels --tokenizer_dir <bpe-dir>
python data_prepro_clean.py --mode binarize_cls_labels --input_dir <processed-data-dir> --output_dir <processed-data-dir>/data_bin --tokenizer_dir <bpe-dir>
Launch training jobs using scripts scripts/launch_multitask_*.py
.
To evaluate the QUALS of summaries (e.g. test.target) given original input (e.g. test.source), we execute the following steps in the preprocess
sub-directory.
python evaluate_hypo.py --mode convert_hypo_to_json --base_dir <processed-data-dir> --sub_dir <any-sub-directory-to-data> --split test --pattern .target
python sm_inference_asum.py --task gen_qa --base_dir <processed-data-dir> --source_dir <any-sub-directory-to-data> --output_dir <output-dir> --num_workers <num-of-gpus> --bsz 5 --beam 60 --max_len 60 --min_len 8 --checkpoint_dir <QAGen-model-dir> --ckp_file checkpoint2.pt --bin_dir <processed-data-dir>/data_bin --diverse_beam_groups 60 --diverse_beam_strength 0.5 --batch_lines True --input_file test.target.hypo --return_token_scores True
Here, we use diverse beam search to generate 60 question-answer pairs for each summary. The batch_lines
option is set to True
to batch bsz
input summaries together for efficient generation. The QAGen model is trained by fine-tuning BART on the SQuAD and NewsQA datasets by concatenating the question-answer pairs using a separator.
To train the QAGen model, place the dev-v1.1.json
and train-v1.1.json
of SQuAD and the combined-newsqa-data-v1.json
of the NewsQA under <squad-newsqa-dir>
. The following command generates the binarized input for fine-tuning BART using Fairseq.
python data_prepro_clean.py --mode newsqa_squad_prepro --input_dir <squad-newsqa-dir> --output_dir <squad-newsqa-dir>
python evaluate_hypo.py --mode filter_qas_dataset_lm_score --base_dir <processed-data-dir> --sub_dir <any-sub-directory-to-qas> --pattern test.target.hypo.beam60.qas
python sm_inference_asum.py --task qa_eval --base_dir <processed-data-dir> --output_dir <output-dir> --num_workers <num-of-gpus> --bsz 30 --checkpoint_dir <QAGen-model-dir> --ckp_file checkpoint2.pt --bin_dir <processed-data-dir>/data_bin --qas_dir <sub-directory-to-qas-filtered> --source_file test.source --target_file test.target --input_file test.target.qas_filtered --prepend_target False
python evaluate_hypo.py --mode compute_hypos_lm_score --base_dir <processed-data-dir> --sub_dir <sub-directory-to-qas-filtered> --pattern test.*.source_eval_noprepend
To use QUALS to improve factual consistency of the summarization model using the CONSEQ algorithm, we follow the steps:
- Obtain the MLE summarization baseline by fine-tuning the BART model. Note that in the ACL paper, we used the corruption-filtered CNNDM and XSUM datasets (
filter_level=1
). - Use the MLE summarization model to sample summaries on the training data.
- Evaluate the QUALS for the generated summaries as well as the ground truth summaries of the training data.
- Form the positive and negative sets for contrastive learning.
- Fine-tune the MLE summarization model using the positive and negative examples. Example scripts for launching training jobs locally or on Sagemaker are
scripts/launch_sagemaker_unlikelihood_cnndm.py
andscripts/launch_sagemaker_unlikelihood_xsum.py
.
We provide an example script preprocess/run_generate_unlikelihood_train_xsum.sh
to illustrate steps 2-4.
Note that
- To avoid running for a long time and encountering OOM errors and then restarting the whole process, we split the input files into smaller ones. We do this by splitting the source file by line (e.g. each sub-file has 10000 lines):
split -l 10000 train.source train.source.split
- The script has to make repeated calls of
python sm_inference_asum.py --task gen_qa
to generate question-ansewr pairs, as many times as there are sub-files as a result of line splits. The python function automatically checks which sub-files have been processed (based on output files) so it always processes the next available sub-file. If all sub-files have been processed, it will simply do nothing so it's safe if it's called more times than there are available sub-files. - Similarly,
sm_inference_asum.py --task qa_eval
needs to be repeated called to cover all sub-files. - The speed of question-answer pairs generation depends on the batch size setting. Depending on the summary file and the
batch_lines
setting, batching is handled differently. If the summary file contains only a single summary per input document,batch_lines
should be set toTrue
andbsz
number of input lines are batched together as input to the QAGen model. If the summary file contains multiple summaries per input document,batch_lines
should be set toFalse
and the batching is done usingbsz
within each input example (line). For example, if there are 6 summaries per line in the summary file, we should setbatch_lines
toFalse
; settingbsz
to 7 will batch all the 7 summaries in a line together, which gives the best speed. (setting it higher won't improve speed since we do not do batching over different lines of input asbatch_lines
isFalse
). On CNN-DM,bsz
of 7 would sometimes result in OOM errors with 16G GPU memory so I use 3 or 4; on XSUM, it is safe to use 7. num_workers
should be the number of GPUs available on the machine. The lines in each input files will be distributed per GPU.- Finally, run the following to concatenate the QUALS scores from the sub-files:
cat *.quals > train.source.source_eval_noprepend.quals
@inproceedings{nan-etal-2021-entity,
title = "Entity-level Factual Consistency of Abstractive Text Summarization",
author = "Nan, Feng and
Nallapati, Ramesh and
Wang, Zhiguo and
Nogueira dos Santos, Cicero and
Zhu, Henghui and
Zhang, Dejiao and
McKeown, Kathleen and
Xiang, Bing",
booktitle = "Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume",
month = apr,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2021.eacl-main.235",
pages = "2727--2733",
}
@inproceedings{nan-etal-2021-improving,
title = {Improving Factual Consistency of Abstractive Summarization via Question Answering},
author = "Nan, Feng and
Nogueira dos Santos, Cicero and
Zhu, Henghui and
Ng, Patrick and
McKeown, Kathleen and
Nallapati, Ramesh and
Zhang, Dejiao and
Wang, Zhiguo and
Arnold, Andrew and
Xiang, Bing",
booktitle = {Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP)},
address = {Virtual},
month = {August},
url = {https://arxiv.org/abs/2105.04623},
year = {2021}
}