Skip to main content
Python Software Foundation 20th Year Anniversary Fundraiser  Donate today!

Toolkit for summarization evaluation

Project description

Summarization Repository

Authors: Alex Fabbri*, Wojciech Kryściński*, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev

This project is a collaboration work between Yale LILY Lab and Salesforce Research.

LILY Logo             Salesforce Logo

* - Equal contributions from authors

Table of Contents

  1. Updates
  2. Data
  3. Evaluation Toolkit
  4. Citation
  5. Get Involved


04/19/2020 - Updated the human annotation file to include all models from paper and metric scores.
04/19/2020 - SummEval is now pip-installable! Check out the pypi page.
04/09/2020 - Please see this comment with code for computing system-level metric correlations!
11/12/2020 - Added the reference-less BLANC and SUPERT metrics!
7/16/2020 - Initial commit! :)


As part of this release, we share summaries generated by recent summarization model trained on the CNN/DailyMail dataset here.
We also share human annotations, collected from both crowdsource workers and experts here.

Both datasets are shared WITHOUT the source articles that were used to generate the summaries.
To recreate the full dataset please follow the instructions listed here.

Model Outputs

Model Paper Outputs Type
M0 Lead-3 Baseline Link Extractive
M1 Neural Document Summarization by Jointly Learning to Score and Select Sentences Link Extractive
M2 BANDITSUM: Extractive Summarization as a Contextual Bandit Link Extractive
M3 Neural Latent Extractive Document Summarization Link Extractive
M4 Ranking Sentences for Extractive Summarization with Reinforcement Learning Link Extractive
M5 Learning to Extract Coherent Summary via Deep Reinforcement Learning Link Extractive
M6 Neural Extractive Text Summarization with Syntactic Compression Link Extractive
M7 STRASS: A Light and Effective Method for Extractive Summarization Based on Sentence Embeddings Link Extractive
M8 Get To The Point: Summarization with Pointer-Generator Networks Link Abstractive
M9 Fast Abstractive Summarization with Reinforce-Selected Sentence Rewriting Link Abstractive
M10 Bottom-Up Abstractive Summarization Link Abstractive
M11 Improving Abstraction in Text Summarization Link Abstractive
M12 A Unified Model for Extractive and Abstractive Summarization using Inconsistency Loss Link Abstractive
M13 Multi-Reward Reinforced Summarization with Saliency and Entailment Link Abstractive
M14 Soft Layer-Specific Multi-Task Summarization with Entailment and Question Generation Link Abstractive
M15 Closed-Book Training to Improve Summarization Encoder Memory Link Abstractive
M16 An Entity-Driven Framework for Abstractive Summarization Link Abstractive
M17 Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer Link Abstractive
M18 Better Rewards Yield Better Summaries: Learning to Summarise Without References Link Abstractive
M19 Text Summarization with Pretrained Encoders Link Abstractive
M20 Fine-Tuning GPT-2 from Human Preferences Link Abstractive
M21 Unified Language Model Pre-training for Natural Language Understanding and Generation Link Abstractive
M22 BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension Link Abstractive
M23 PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization Link Abstractive


All model outputs were obtained from the original authors of the models and shared with their consent.
When using any of the model outputs, please also cite the original paper.

Human annotations

Human annotations of model generated summaries can be found here.

The annotations include summaries generated by 16 models from 100 source news articles (1600 examples in total).
Each of the summaries was annotated by 5 indepedent crowdsource workers and 3 independent experts (8 annotations in total).
Summaries were evaluated across 4 dimensions: coherence, consistency, fluency, relevance.
Each source news article comes with the original reference from the CNN/DailyMail dataset and 10 additional crowdsources reference summaries.

Data preparation

Both model generated outputs and human annotated data require pairing with the original CNN/DailyMail articles.

To recreate the datasets follow the instructions:

  1. Download CNN Stories and Daily Mail Stories from
  2. Create a cnndm directory and unpack downloaded files into the directory
  3. Download and unpack model outputs or human annotations.
  4. Run the script to pair the data with original articles

Example call for model outputs:

python3 data_processing/ --model_outputs <file-with-data-annotations> --story_files <dir-with-stories>

Example call for human annotations:

python3 data_processing/ --data_annotations <file-with-data-annotations> --story_files <dir-with-stories>

Evaluation Toolkit

We provide a toolkit for summarization evaluation to unify metrics and promote robust comparison of summarization systems. The toolkit contains popular and recent metrics for summarization as well as several machine translation metrics.


Below are the metrics included in the tookit, followed by the associated paper and code used within the toolkit: |Metric|Paper|Code| |-|-|-| |ROUGE|ROUGE: A Package for Automatic Evaluation of Summaries|Link| |ROUGE-we|Better Summarization Evaluation with Word Embeddings for ROUGE|Link| |MoverScore|MoverScore: Text Generation Evaluating with Contextualized Embeddings and Earth Mover Distance|Link| |BertScore|BertScore: Evaluating Text Generation with BERT|Link| |Sentence Mover's Similarity|Sentence Mover’s Similarity: Automatic Evaluation for Multi-Sentence Texts|Link| |SummaQA|Answers Unite! Unsupervised Metrics for Reinforced Summarization Models|Link| |BLANC|Fill in the BLANC: Human-free quality estimation of document summaries|Link| |SUPERT|SUPERT: Towards New Frontiers in Unsupervised Evaluation Metrics for Multi-Document Summarization|Link| |METEOR|METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments |Link| |S3|Learning to Score System Summaries for Better Content Selection Evaluation|Link| |Misc. statistics
(extractiveness, novel n-grams, repetition, length)|Newsroom: A Dataset of 1.3 Million Summaries with Diverse Extractive Strategies| Link| |Syntactic Evaluation|Automatic Analysis of Syntactic Complexity in Second Language writing|Link| |CIDer|CIDEr: Consensus-based Image Description Evaluation|Link| |CHRF|CHRF++: words helping character n-grams|Link| |BLEU|BLEU: a Method for Automatic Evaluation of Machine Translation|Link|


You can install summ_eval via pip:

pip install summ-eval

You can also install summ_eval from source:

git clone
cd evaluation
pip install -e .

You can test your installation (assuming you're in the ./summ_eval folder) and get familiar with the library through tests/

python -m unittest discover

Command-line interface

We provide a command-line interface calc-scores which makes use of gin config files to set metric parameters.


Run ROUGE on given source and target files and write to rouge.jsonl, analogous to files2rouge.

calc-scores --config-file=examples/basic.config --metrics "rouge" --summ-file summ_eval/1.summ --ref-file summ_eval/1.ref --output-file rouge.jsonl --eos " . " --aggregate True

NOTE: if you're seeing slow-ish startup time, try commenting out the metrics you're not using in the config; otherwise this will load all modules.

Run ROUGE and BertScore on a jsonl file which contains "reference" and "summary" keys and write to output.jsonl.

calc-scores --config-file=examples/basic.config --metrics "rouge, bert_score" --jsonl-file data.jsonl --output-file rouge_bertscore.jsonl

For a full list of options, please run:

calc-scores --help

For use in scripts

If you want to use the evaluation metrics as part of other scripts, we have you covered!

from summ_eval.rouge_metric import RougeMetric
rouge = RougeMetric()

Evaluate on a batch

summaries = ["This is one summary", "This is another summary"]
references = ["This is one reference", "This is another"]

rouge_dict = rouge.evaluate_batch(summaries, references)

Evaluate on a single example

rouge_dict = rouge.evaluate_example(summaries[0], references[0])

Evaluate with multiple references

Currently the command-line tool does not use multiple references for simplicity. Each metric has a supports_multi_ref property to tell you if it supports multiple references.

print(rouge.supports_multi_ref) # True
multi_references = [["This is ref 1 for summ 1", "This is ref 2 for summ 1"], ["This is ref 1 for summ 2", "This is ref 2 for summ 2"]]
rouge_dict = rouge.evaluate_batch(summaries, multi_references)


  title={SummEval: Re-evaluating Summarization Evaluation},
  author={Fabbri, Alexander R and Kry{\'s}ci{\'n}ski, Wojciech and McCann, Bryan and Xiong, Caiming and Socher, Richard and Radev, Dragomir},
  journal={arXiv preprint arXiv:2007.12626},

Get Involved

Please create a GitHub issue if you have any questions, suggestions, requests or bug-reports. We welcome PRs!

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for summ-eval, version 0.24
Filename, size File type Python version Upload date Hashes
Filename, size summ_eval-0.24-py3-none-any.whl (62.5 MB) File type Wheel Python version py3 Upload date Hashes View
Filename, size summ_eval-0.24.tar.gz (62.3 MB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page