A Shared Task at EMNLP 2020 that focuses on generation of long summaries for scientific documents. LongSumm is one of three shared tasks conducted as part of: 1st Workshop on Scholarly Document Processing
Most of the work on scientific document summarization focuses on generating relatively short summaries. Such a length constraint might be appropriate when summarizing news articles but it is less adequate for scientific work. In fact, such a short summary resembles an abstract and cannot cover all the salient information conveyed in a given scientific text. Writing longer summaries requires expertise and a deep understanding in a scientific domain, as can be found in some researchers blogs.
To address this point, the LongSumm task opted to leverage blog posts created by researchers in the NLP and Machine learning communities that summarize scientific articles and use these posts as reference summaries.
The corpus for this task includes a training set that consists of 1705 extractive summaries, and 531 abstractive summaries of NLP and Machine Learning scientific papers. The extractive summaries are based on video talks from associated conferences (Lev et al. 2019 TalkSumm) while the abstractive summaries are blog posts created by NLP and ML researchers. In addition, we created a test set of abstractive summaries for testing submissions. Each submission is judged against one reference summary (gold summary) using ROUGE and should not exceed 600 words.
The 1st Workshop on Scholarly Document Processing will include two additional shared tasks:
- Scisumm - focuses on automatic paper summarization on a new corpus of research papers in Computational Linguistics (CL) domain
- LaySumm - focuses on enabling systems to automatically generate lay summaries. A lay summary explains, succinctly and without using technical jargon, what the overall scope, goal and potential impact of a scientific paper is.
You are invited to participate in the LongSumm Shared Task at SDP@EMNLP 2020. This repository contains the dataset and instructions on how to participate in the task.
The training data is composed of abstractive and extractive summaries.
The abstractive summaries are from different domains of CS including ML, NLP, AI, vision, storage, etc.
The training data contains around 700 abstractive summaries that can be found at data/abstractive/cluster. The folder contains clusters of summaries with length varying between 100-1500 words. Each sub-folder clusters into bins of size 100 words. (i.e., summary of 541 words will appear in the corresponding cluster of 500-600). We used the Python NLTK library to count the number of words and to segment summary text into sentences.
The format of a summary is a JSON file with the following entries:
Entry | Description |
---|---|
id | Record id (unique) |
blog_id | The id of the blog |
summary | An array of the sentences of the summary |
author_id | The id of the author |
pdf_url | The link to the original paper |
author_full_name | The author full name |
source_website | the website in which the original blog appears |
Example:
{
"id": "79792577",
"blog_id": "4d803bc021f579d4aa3b24cec5b994",
"summary": [
"Task of translating natural language queries into regular expressions ...",
"Proposes a methodology for collecting a large corpus of regular expressions ...",
"Reports performance gain of 19.6% over state-of-the-art models.",
"Architecture LSTM based sequence to sequence neural network (with attention) Six layers ...",
"Attention over encoder layer.",
"...."
],
"author_id": "shugan",
"pdf_url": "http://arxiv.org/pdf/1608.03000v1",
"author_full_name": "Shagun Sodhani",
"source_website": "https://github.com/shagunsodhani/papers-I-read"
}
Each papers' summary should be linked the corresponding text of the original paper. Due to copyright restrictions will not publish the original papers, here are the suggested steps to fully construct the dataset:
-
Extract PDF - to download the PDF of each paper, one can use the following script : downloader.py. The output of this scripts is the papers PDFs by their IDs, under the out_folder.
Notice - some of the papers may require a subscription (e.g., ACM). If you do not have the permission the script won't be able to download the paper.
The script accepts as input 3 parameters :
clusters_dir
- path to the directory that contains the summariesout_folder
- path to the output directory where you want all the PDFsnum_processes
- the script has an option to run in a multiprocess fashion. Default=1, we recommend to use more in order to decrease the downloading time.
python downloader.py --clusters_dir=/path/to/input/dir/with/clusters --out_folder=/path/to/output/dir/for/PDF --num_processes=3
-
Extract Text of the PDF- given papers in pdf format, we recommend to use science-parse to convert them to structured json files.
At the end of this step, you should have for each summary, a corresponding JSON file of the original text from the paper as extracted by science-parse.
The extractive summaries are based on the TalkSumm (Lev et al. 2019) dataset. The dataset contains 1705 automatically-generated noisy extractive summaries of scientific papers from the NLP and Machine Learning domain based on video talks from associated conferences (e.g., ACL, NAACL, ICML) Summaries can be found under data/extractive/. Each summary provides the top-30 sentences, which are on average around 990 words. The format of each summary file is as follows:
- Each line contains: sentence index (in original paper), sentence score (i.e. duration), then the sentence itself. The fields are tab-separated.
- The order of the sentences is according to their order in the paper.
- Link to the reference paper.
If you wish to create extractive summaries of a paper that doesn't not exist in the dataset, you will need to follow the instructions from: https://github.com/levguy/talksumm
There are 22 papers for the test set, as listed below.
Paper id | Paper title | Paper link |
---|---|---|
1000 | Unsupervised Part-of-Speech Tagging with Bilingual Graph-Based Projections | https://www.aclweb.org/anthology/P11-1061.pdf |
1001 | RNN Fisher Vectors for Action Recognition and Image Annotation | https://arxiv.org/pdf/1512.03958.pdf |
1002 | TALK SUMM: A Dataset and Scalable Annotation Method for Scientific Paper Summarization Based on Conference Talks | https://arxiv.org/pdf/1906.01351.pdf |
1003 | Emotion Detection from Text via Ensemble Classification Using Word Embeddings | https://dl.acm.org/doi/pdf/10.1145/3121050.3121093 |
1004 | Classifying Emotions in Customer Support Dialogues in Social Media | https://www.aclweb.org/anthology/W16-3609.pdf |
1005 | MetAdapt: Meta-Learned Task-Adaptive Architecture for Few-Shot Classification | https://arxiv.org/pdf/1912.00412.pdf |
1006 | Detecting Egregious Conversations between Customers and Virtual Agents | https://www.aclweb.org/anthology/N18-1163.pdf |
1007 | Understanding Convolutional Neural Networks for Text Classification | https://www.aclweb.org/anthology/W18-5408.pdf |
1008 | An Editorial Network for Enhanced Document Summarization | https://www.aclweb.org/anthology/D19-5407.pdf |
1009 | DIMSIM: An Accurate Chinese Phonetic Similarity Algorithm Based on Learned High Dimensional Encoding | https://www.aclweb.org/anthology/K18-1043.pdf |
1010 | Improved Neural Relation Detection for Knowledge Base Question Answering | https://www.aclweb.org/anthology/P17-1053.pdf |
1011 | Interactive Dictionary Expansion using Neural Language Models | http://ceur-ws.org/Vol-2169/paper-02.pdf |
1012 | Interpretable and Globally Optimal Prediction for Textual Grounding using Image Concepts | https://papers.nips.cc/paper/6787-interpretable-and-globally-optimal-prediction-for-textual-grounding-using-image-concepts.pdf |
1013 | Learning Implicit Generative Models by Matching Perceptual Features | https://arxiv.org/pdf/1904.02762.pdf |
1014 | Scalable Demand-Aware Recommendation | http://papers.nips.cc/paper/6835-scalable-demand-aware-recommendation.pdf |
1015 | Neural Response Generation for Customer Service based on Personality Traits | https://www.aclweb.org/anthology/W17-3541.pdf |
1016 | A Low Power, High Throughput, Fully Event-Based Stereo System | https://openaccess.thecvf.com/content_cvpr_2018/CameraReady/3791.pdf |
1017 | Characterization and Learning of Causal Graphs with Latent Variables from Soft Interventions | https://papers.nips.cc/paper/9581-characterization-and-learning-of-causal-graphs-with-latent-variables-from-soft-interventions.pdf |
1018 | Complex Program Induction for Querying Knowledge Bases in the Absence of Gold Programs | https://www.aclweb.org/anthology/Q19-1012.pdf |
1019 | Unsupervised Dual-Cascade Learning with Pseudo-Feedback Distillation for Query-based Extractive Summarization | https://arxiv.org/pdf/1811.00436 |
1020 | High quality, lightweight and adaptable TTS using LPCNet | https://arxiv.org/pdf/1905.00590 |
1021 | Sobolev Independence Criterion | https://arxiv.org/pdf/1910.14212.pdf |
The intrinsic evaluation will be done by ROUGE, using ROUGE-1, -2, -L metrics. In addition, a randomly selected subset of the summaries will undergo human evaluation.
The submission should be a single json file containing all summaries, following the format:
{
"paper_id_1":"summary of paper 1",
"paper_id_2":"summary of the paper 2"
}
https://github.com/guyfe/LongSumm/blob/master/scripts/evaluation_script.py
In order to submit you will need to follow these steps:
- Create an IBM account at ibm.com (https://tinyurl.com/ydcd6hjg). Please use the email account that you registered to the task.
- Login to the AI Leaderboard (https://aieval.draco.res.ibm.com/) with your IBM account
AI Leaderboard instructions:
- Choose "Participate"
- Choose "Participant Teams" and create a new participant team - use a meaningful name for your group as this is the name that will appear in the leaderboard.
- Go to "All Challenges" and find our task "Long Scientific Document Summarization" , and click on "View Details"
- To submit go the the "Participate" tab, select a team, click on "Next" and accept the Terms & Condisiton. (this step is done only once). In case that you are not able to click on "Next" please press refresh that page (Ctrl+Shift+R or Cmnd+Shift+r)
- Go to the "Submit" tab, there you can upload the json file, describe the submission, and press "Submit".
- To view your submission(s) go to the "My Submissions". There you can see all your submissions, their status (Finished/Failed), links to some logs, and results. Finally in case that you want your submission to appear in the leaderboard you will need to check "Show on leaderboard".
- Finally, in order to see the leaderboard go to the "Leaderboard" tab.
- Firefox and Chrome browsers are supported
- In any case that you seems not to see submission/results on leaderboard press refresh that page (Ctrl+Shift+R or Cmnd+Shift+r)
- You can submit up to 25 runs.
You should only submit summaries that are part of the test data. Please do not submit any confidential or personal information. Please see the IBM Terms of use (https://www.ibm.com/legal)
We would like to thank the following blog authors and to ShortScience.org who genereously allowed us to share the content as part of this dataset.
- Shagun Sodhani https://github.com/shagunsodhani/papers-I-read
- Patrick Emami https://pemami4911.github.io/index.html
- Adrian Colyer https://blog.acolyer.org/about/
- Alexander Jung https://github.com/aleju/papers
- Joseph Paul Cohen https://www.shortscience.org/user?name=joecohen
- Hugo Larochelle https://www.shortscience.org/user?name=hlarochelle
- Elvis Saravia https://github.com/dair-ai/nlp_paper_summaries
- Abstractive summaries:
- ShortScience.org summaries are released under Attribution-NonCommercial-ShareAlike 4.0
- all other summaries are released under the CDLA-Sharing license https://cdla.io/sharing-1-0/
- Extractive summaries - released under Attribution-NonCommercial-ShareAlike 4.0.
The data was copied from the above mentioned blogs as-is. IBM is not responsible for the content of the data, nor for any claim related to the data (including claims related to alleged intellectual property or privacy breach).
For further information about this dataset please contact the organizers of the shared task: