CC30k: A Citation Contexts Dataset for Reproducibility-Oriented Sentiment Analysis

Paper | Code

The CC30k dataset consists of labeled citation contexts obtained through crowdsourcing. Each context is labeled by three independent workers. This README describes the structure and columns of the dataset.

Dataset Description

The CC30k dataset is unique in its focus on reproducibility-oriented sentiments (ROS) within scientific literature. This introduces a novel approach to studying computational reproducibility by leveraging citation contexts, which are textual fragments in scientific papers that reference prior work. This dataset comprises 30,734 labeled citation contexts from scientific literature published at AI venues, each annotated with one of three ROS labels: Positive, Negative, or Neutral. These labels reflect the cited work's perceived reproducibility. The dataset contains ROS labeled contexts along with metadata about the workers, reproducibility study, related original paper, citing paper, the final aggregated labels and the label type. The columns in the dataset are detailed in the table below:

Column Name	Description
input_index	Unique ID for each citation context.
input_context	Citation context that workers are asked to label.
input_file_key	Identifier linking the context to a rep-study.
input_first_author	Name or identifier of the first author of the cited paper.
worker_id_w1	Unique ID of the first worker who labeled this citation context.
work_time_in_seconds_w1	Time (in seconds) the first worker took to label the citation context.
worker_id_w2	Unique ID of the second worker who labeled this citation context.
work_time_in_seconds_w2	Time (in seconds) the second worker took to label the citation context.
worker_id_w3	Unique ID of the third worker who labeled this citation context.
work_time_in_seconds_w3	Time (in seconds) the third worker took to label the citation context.
label_w1	Label assigned by the first worker.
label_w2	Label assigned by the second worker.
label_w3	Label assigned by the third worker.
batch	Batch number for the posted Mechanical Turk job.
majority_vote	Final label based on the majority vote among workers’ labels (reproducibility-oriented sentiment: Positive, Negative, or Neutral).
majority_agreement	Indicates how many of the three workers agreed on the final majority vote.
rs_doi	Digital Object Identifier (DOI) of the reproducibility study paper.
rs_title	Title of the reproducibility study paper.
rs_authors	List of authors of the reproducibility study paper.
rs_year	Publication year of the reproducibility study paper.
rs_venue	Venue (conference or journal) where the reproducibility study was published.
rs_selected_claims	Number of claims selected from the original paper for reproducibility study (by manual inspection).
rs_reproduced_claims	Number of selected claims that were successfully reproduced (by manual inspection).
reproducibility	Final reproducibility label assigned to the original paper by manual inspection (reproducible, not-reproducible, partially-reproducible [if 0 < rs_reproduced_claims < rs_selected_claims]).
org_doi	DOI of the original (cited) paper that was assessed for reproducibility.
org_title	Title of the original (cited) paper.
org_authors	List of authors of the original (cited) paper.
org_year	Publication year of the original (cited) paper.
org_venue	Venue where the original (cited) paper was published.
org_paper_url	URL to access the original (cited) paper.
org_citations	Number of citations received by the original (cited) paper.
org_s2ga_id	Semantic Scholar Graph API ID of the original (cited) paper.
citing_doi	DOI of the citing paper that cited the original (cited) paper.
citing_year	Publication year of the citing paper.
citing_venue	Venue where the citing paper was published.
citing_title	Title of the citing paper.
citing_authors	List of authors of the citing paper.
citing_s2ga_id	Semantic Scholar Graph API ID of the citing paper.
label_type	Label source: crowdsourced or augmented_human_validated or augmented_machine_labeled.

Jupyter Notebook Descriptions

The GitHub repository's notebooks directory contains the following Jupyter notebooks, which were used to produce and analyze the dataset:

• R001_AWS_Labelling_Dataset_Preprocessing_Mturk.ipynb
  • Used to pre-process data for Mechanical Turk (MTurk) labeling.
• R001_AWS_MTurk_API.ipynb
  • Used to communicate with MTurk workers.
• R001_AWS_MTurk_process_results.ipynb
  • Used to process crowdsourced results from MTurk.
• R001_Extend_CC25k_Dataset.ipynb
  • Used to extend the crowdsourced labels with newly augmented ROS: Negative contexts.
• R_001_Creating_the_RS_superset.ipynb
  • Used to collect the original and reproducibility studies.
• R_001_Extract_Citing_Paper_Details.ipynb
  • Used to collect citing paper details and contexts using the Semantic Scholar Graph API (S2GA).
• R001_MTurk_Sentiment_Analysis_5_models.ipynb
  • Generates the performance measures for the selected five open-source multiclass sentiment analysis models.

Citation

@misc{obadage2025cc30kcitationcontextsdataset,
      title={CC30k: A Citation Contexts Dataset for Reproducibility-Oriented Sentiment Analysis}, 
      author={Rochana R. Obadage and Sarah M. Rajtmajer and Jian Wu},
      year={2025},
      eprint={2511.07790},
      archivePrefix={arXiv},
      primaryClass={cs.DL},
      url={https://arxiv.org/abs/2511.07790}, 
}