Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


** LaySumm is NOT covered by CC BY 4.0 licence. Please do not email us about Elsevier's LaySumm. We are unable to respond**

Shield: CC BY 4.0

This work is licensed under a Creative Commons Attribution 4.0 International License EXCEPT for the following files which are closed source under strict copyright laws enforced by Elsevier labs. We hold no accountability for these:

CC BY 4.0

(scisumm-corpus @

This package contains a release of training and test topics to aid in the development of computational linguistics summarization systems.


The CL-SciSumm Shared Task is run off the CL-SciSumm corpus and is composed of
three sub-tasks in automatic research paper summarization on a new corpus of research papers. A training corpus with summaries for one thousand forty topics and forty topics for citance to reference span id (or provenance identification) tasks has been released. A test corpus of twenty topics is held-out as a blind test-set. The topics comprise of ACL Computational Linguistics and Natural Language Processing research papers, and their citing papers and three output summaries each. The three output summaries comprise: the traditional authors' summary of the paper (the abstract), the community summary (the collection of citation sentences ‘citances’) and a human summary written by a trained annotator. Within the corpus, each citance is also mapped to its referenced text in the reference paper and tagged with the information facet it represents.

The manually annotated training set of 40 articles (Tasks 1a and b) and citing papers, human written summaries (1040 documents) for them and a further 1000 document corpus (ScisummNet), an auto-annotated noisy dataset with several thousands of article-citing paper papers (to aid in ' training deep learning models) are readily available for download and can be used by participants. This data can be found in /data/Training-Set-2019/Task1/From-Training-Set-2018 and /data/Training-Set-2019/Task2/From-Training-Set-2018

The last edition of CL-SciSumm was CL-SciSumm 2020. The gold test data used for 2020, 2019, 2018 are now available in public in this repo. You are welcome to use it for your evaluations and paper submissions to any conference / journal or your theses.

In 2020, we did not add any new training data.

In 2019 we had introduced 1000 document sets that were automatically annotated to be used as training data. This training data was generated following Nomoto,2018. This data can be found in /data/Training-Set-2019/Task1/From-ScisummNet-2019. Note that the auto-annotated data is available only for Task 1a. No discourse facet is provided for the classification task: Task1b. We recommend you to use the auto-anootated data only for training 'reference span selection' models for Task 1a and use the manually annotated training data from 40 document sets for Task1b.

Further, for Task 2 one thousand summaries that were released as part of the SciSummNet (Yasunaga et al., 2019) have been included as human summaries to train on. This data can be found in /data/Training-Set-2019/Task2/From-ScisummNet-2019

The test set of 20 articles is available in /data/Test-Set-2018. This is a blind test set, that is, the ground truth is withheld. The system outputs from the test set should be submitted to the task organizers, for the collation of the final results to be presented at the workshop.

For more details, see the Contents Section at the bottom of this Readme. To know how this corpus was constructed, please see ./docs/corpusconstruction.txt

Last editions proceedings:

If you use the data and publish please let us know and cite our CL-SciSumm 2019 task overview paper:

title={Overview and Results: CL-SciSumm Shared Task 2019},<br>
author={Chandrasekaran, Muthu Kumar and Yasunaga, Michihiro and Radev, Dragomir and Freitag, Dayne and Kan, Min-Yen},<br>
booktitle={In Proceedings of Joint Workshop on Bibliometric-enhanced Information Retrieval and NLP for Digital Libraries (BIRNDL 2019)},<br>


CL-SciSumm ran as a shared task at EMNLP 2020, SIGIR 2019, 2018, 2017, JCDL 2016 and the Pilot Task conducted as a part of the BiomedSumm Track at the Text Analysis Conference 2014 (TAC 2014).

The task is on automatic paper summarization in the Computational Linguistics (CL) domain. The output summaries are of two types: faceted summaries of the traditional self-summary (the abstract) and the community summary (the collection of citation sentences ‘citances’). We also group the citances by the facets of the text that they refer to.

The task is defined as follows:

Given: A topic consisting of a Reference Paper (RP) and upto 10 Citing Papers (CPs) that all contain citations to the RP. In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.

  • Task 1a: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance. These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).
  • Task 1b: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.
  • Task 2 (optional bonus task): Finally, generate a structured summary of the RP from the cited text spans of the RP. The length of the summary should not exceed 250 words.

Evaluation: Task 1 is scored by overlap of text spans measured by number of sentences in the system output vs gold standard. Task 2 is scored using the ROUGE family of metrics between i) the system output and the gold standard summary fromt the reference spans ii) the system output and the asbtract of the reference paper.


This is the open repository for the Scientific Document Summarization Corpus and Annotations contributed to the public by the Web IR / NLP Group at @ the National University of Singapore (WING-NUS) with generous support from Microsoft Research Asia.


This file.


Frequently asked questions on the 2018 shared task including updates to the corpus, annotation format from the previous edition.


README files for the previous editionS of the shared task hosted at BIRNDL@SIGIR 2018, BIRNDL@SIGIR 2017, BIRNDL@JCDL 2016 and TAC2014.


A readme detailing the rules and steps followed to create the document corpus by randomly sampling documents from the ACL Anthology corpus and selecting their citing papers.


Describes the naming convention followed to identify annotation files for each training topic in ./data/???-????_TRAIN/Annotation/


Rules followed to resolve difficult cases in annotation. It can serve as a synopsis of the larger annotation guidelines. For the detailed annotation guidelines, please refer to the details hosted at


References for each of the papers for each of the topics, one file per topic.


Directories containing the Documents, Summaries, and Annotations for each topic, one directory per Topic ID.


This directory contains the 10 source documents for the topic (1 RP and upto 10 CPs), one file per paper, in the original pdf format.


This directory contains the source document for the RP of the topic in XML format in UTF-8 character encoding. The file corresponds to the similarly named pdf file in Documents_PDF/. All annotations and offsets for the topic are with respect to the xml files in this directory. All the files were created from the pdf file using Adobe Acrobat.
Note that there were OCR errors in reading several of the files, and the annotators often had to manually edit the converted txt files. Research groups using are free to use alternative parsing tools on the pdfs provided, if they are found to perform better.


This directory contains the source document for the CPs of the topic in xml format in UTF-8 character encoding. Each file corresponds to the similarly named pdf file above.


This directory contains the annotation files for the topic, from 3 different annotators.
Please DO NOT use older annotations; only use .annv3.txt for the 2016 Shared Task.


The summary task (Task 2) is an optional, "bonus" task which participants may want to attempt. This directory contains the two kinds of summaries - i. the abstract, and ii.human written summaries of the reference paper.


Given a reference paper (RP) and 10 or more citing papers (CPs), annotators from the University of Hyderbad were instructed to find citations to the RP in the CPs. Annotators followed instructions in SciSumm-annotation-guidelines.pdf to mark the Citation Text, Citation Marker, Reference Text, and Discourse Facet for each citation of the RP found in the CP.

Organisers' Contacts

Please open github issues for further information or to report a bug or a fix for the corpus.

Contacts for maintainers of the corpus:

We provide links for Laysumm here and are not associated with it:

README for Lay Summarization Task (LaySumm 2020)

Task Description and a sample dataset can be found at: here in this Github repo.


This README was updated from README2020 by Muthu Kumar Chandrasekaran in 2021. For revision information, check source code control logs.


Scientific Document Summarization Corpus and Annotations from the WING NUS group.







No releases published


No packages published