Datasets:
dataset_info:
features:
- name: id
dtype: string
- name: title
dtype: string
- name: abstract
dtype: string
- name: type
dtype: string
splits:
- name: train
num_bytes: 65753436
num_examples: 44500
download_size: 39508470
dataset_size: 65753436
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: apache-2.0
language:
- en
- es
- fr
- de
- it
- pt
size_categories:
- 10K<n<100K
Dataset Description
The unlabelled-sti-corpus is a diverse dataset designed for developing information extraction datasets (i.e. text classification or NER) for Science, Technology, and Innovation (STI) records. The corpus contains approximately 35,000 records sourced from four major repositories:
- 22,500 publications from OpenAlex
- 10,000 European research projects from CORDIS
- 5,000 regional projects from Interreg and Kohesio
- 7,000 patents from Lens.org
The dataset is stratified across disciplines and field categories to ensure representation of a wide range of topics and domains. It is intended for labelling tasks that can support a variety of downstream applications in STI research, including extracting entities, relations, and other structured information from textual data. The dataset is unannotated and intended for labelling tasks, meaning users must prepare their own labelled subsets for specific applications.
Dataset Creation
Curation Rationale
The dataset was created to address the need for a diverse and representative corpus for information extraction in STI domains. By including records from multiple repositories and stratifying by discipline, the corpus provides a robust foundation for developing models applicable to a wide range of STI applications.
Source Data
The dataset aggregates records from:
- OpenAlex: A comprehensive open catalogue of scholarly works.
- CORDIS: The EU's primary repository for research project information.
- Interreg: A program fostering interregional cooperation in Europe.
- Kohesio: A database of regional development projects.
- Patstat: Lens.org contains bibliographical and legal event patent data from different jurisdictions and patent offices
Stratification by field categories was conducted to ensure that records from various disciplines (e.g., engineering, life sciences, social sciences) are represented.