Dataset Card for scitechnews

Dataset Summary

The SciTechNews dataset consists of scientific papers paired with their corresponding press release snippet mined from ACM TechNews. ACM TechNews is a news aggregator that provides regular news digests about scientific achieve- ments and technology in the areas of Computer Science, Engineering, Astrophysics, Biology, and others.

Supported Tasks and Leaderboards

This dataset was curated for the task of Science Journalism, a text-to-text task where the input is a scientific article and the output is a press release summary. However, this release also include additional information of the press release and of the scientific article, such as press release article body, title, authors' names and affiliations.

The science juornalism leaderboard is found here.

Languages

English

Dataset Structure

Data Fields

{
    "id": String,          			 # unique ID
    "pr-title": String,    		 	 # Title as found in the ACMTECHNEWS website
    "pr-article": String,  			 # Press release article
    "pr-summary": String,  		 	 # Press release summary
    "sc-title": String,    		 	 # Title of scientific article
    "sc-abstract": String, 		 	 # Abstract of scientific article
    "sc-article": String,  			 # Concatenated abstract and sections of the scientific article
    "sc-sections": List[String], 	 # List of sections in the scientific article
    "sc-section_names": List[String] # List of section names
    "sc-authors": List[String] 		 # list of authors' name and affiliations, in the format '<name> | <affil>'
}

Paragraphs in the press release articles (pr-article) and sections of the scientific article (sc-sections) are separated by \n. Data is not sentence or word tokenized.
Note that field sc-article includes the article's abstract as well as its sections.

Example Instance

{
  "id": 37,
  "pr-title": "What's in a Developer's Name?",
  "pr-article": "In one of the most memorable speeches from William Shakespeare's play, Romeo and Juliet , Juliet ponders, \" What's in a name? That which...",
  "pr-summary": ""Researchers at the University of Waterloo's Cheriton School of Computer Science in Canada found a software developer's perceived race and ethnicity,...",
  "sc-title": On the Relationship Between the Developer's Perceptible Race and Ethnicity and the Evaluation of Contributions in OSS",
  "sc-abstract": "Context: Open Source Software (OSS) projects are typically the result of collective efforts performed by developers with different backgrounds...",
  "sc-articles": "Context: Open Source Software (OSS) projects are typically the result of .... In any line of work, diversity regarding race, gender, personality...",
  "sc-sections": ["In any line of work, diversity regarding race, gender, personality...","To what extent is the submitter's perceptible race and ethnicity related to...",...],
  "sc-section_names": ["INTRODUCTION", "RQ1:", "RQ2:", "RELATED WORK",...],
  "sc-authors": ["Reza Nadri | Cheriton School of Computer Science, University of Waterloo", "Gema Rodriguez Perez | Cheriton School of ...",...]
}

Data Splits

Number of instances in train/valid/test are 26,368/1431/1000.
Note that the training set has only press release data (pr-*), however splits validation and test do have all fields.

Dataset Creation

Curation Rationale

Science journalism refers to producing journalistic content that covers topics related to different areas of scientific research. It plays an important role in fostering public understanding of science and its impact. However, the sheer volume of scientific literature makes it challenging for journalists to report on every significant discovery, potentially leaving many overlooked.
We construct a new open-access high-quality dataset for automatic science journalism that covers a wide range of scientific disciplines.

Source Data

Press release snippets are mined from ACM TechNews and their respective scientific articles are mined from reputed open-access journals and conference proceddings.

Initial Data Collection and Normalization

We collect archived TechNews snippets between 1999 and 2021 and link them with their respective press release articles. Then, we parse each news article for links to the scientific article it reports about. We discard samples where we find more than one link to scientific articles in the press release. Finally, the scientific articles are retrieved in PDF format and processed using Grobid. Following collection strategies of previous scientific summarization datasets, section heading names are retrieved, and the article text is divided into sections. We also extract the title and all author names and affiliations.

Who are the source language producers?

All texts in this dataset (titles, summaries, and article bodies) were produced by humans.

Considerations for Using the Data

Social Impact of Dataset

The task of automatic science journalism is intended to support journalists or the researchers themselves in writing high-quality journalistic content more efficiently and coping with information overload. For instance, a journalist could use the summaries generated by our systems as an initial draft and edit it for factual inconsistencies and add context if needed. Although we do not foresee the negative societal impact of the task or the accompanying data itself, we point at the general challenges related to factuality and bias in machine-generated texts, and call the potential users and developers of science journalism applications to exert caution and follow up-to-date ethical policies.

Additional Information

Dataset Curators

• Ronald Cardenas, University of Edinburgh
• Bingsheng Yao, Rensselaer Polytechnic Institute
• Dakuo Wang, Northeastern University
• Yufang Hou, IBM Research Ireland

Citation Information

Provide the BibTex-formatted reference for the dataset. For example:

@article{cardenas2023dont,
      title={'Don't Get Too Technical with Me': A Discourse Structure-Based Framework for Science Journalism}, 
      author={Ronald Cardenas and Bingsheng Yao and Dakuo Wang and Yufang Hou},
      year={2023},
      eprint={2310.15077},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}