Datasets:
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- relation-extraction
size_categories:
- 1K<n<10K
dataset_info:
features:
- name: text
dtype: string
- name: entity1
dtype: string
- name: entity2
dtype: string
- name: relation
dtype: string
- name: prompt_0_shot
dtype: string
- name: prompt_2_shot
dtype: string
- name: prompt_5_shot
dtype: string
splits:
- name: train
num_bytes: 18464886
num_examples: 7200
- name: validation
num_bytes: 2032939
num_examples: 800
- name: test
num_bytes: 6968890
num_examples: 2717
download_size: 11578042
dataset_size: 27466715
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
Dataset Card for Transformed SemEval 2010 Task 8
Dataset Description
Dataset Summary
This dataset is released as part of the paper "Sub-Billion, Super-Frontier: Fine-Tuned Small Language Models Rival Zero-Shot Frontier LLMs on General and Literary Relation Extraction" (Christou & Tsoumakas, 2026) arXiv:2606.22606.
The dataset is a transformed version of the SemEval 2010 Task 8 dataset (SemEvalWorkshop/sem_eval_2010_task_8). The original dataset is a standard benchmark for relation extraction and classification between nominal pairs.
This version preprocesses the original data into a more readily usable format with distinct features for entities and cleaned sentence text. Specifically:
- Entity text spans (, ) are extracted into separate fields.
- Special entity tags are removed from the main sentence text.
- Relation labels are provided as their full text names (e.g., "Cause-Effect(e1,e2)") instead of numerical IDs.
- Placeholder fields for entity types are included (set to "none").
The dataset contains training and testing splits as provided in the original SemEval task.
Supported Tasks and Leaderboards
- Relation Extraction: The primary task is to classify the semantic relationship between the two marked entities (
entity1,entity2) within the giventext. - Few-Shot Learning: The dataset structure is suitable for N-shot evaluation protocols.
- Zero-Shot Learning: The dataset can be used for zero-shot evaluation, typically by leveraging the relation name strings as semantic information (requires further setup to define seen/unseen classes and map inputs to relation name embeddings).
Languages
The text in the dataset is in English (BCP-47: en).
Dataset Structure
Data Instances
A typical example from the dataset looks like this:
{
'entity1': 'television',
'entity2': 'programmes',
'entity1Type': 'none',
'entity2Type': 'none',
'relation': 'Product-Producer(e2,e1)',
'text': 'Most programmes have commercial breaks, but the quality of the television programmes depends on the channel.'
}
Data Fields
The dataset contains the following fields:
entity1: (string) The text content of the first marked entity (head entity, originally marked with<e1>).entity2: (string) The text content of the second marked entity (tail entity, originally marked with<e2>).entity1Type: (string) Placeholder for the type of the first entity. Always set to"none"in this version as the original dataset does not provide explicit types.entity2Type: (string) Placeholder for the type of the second entity. Always set to"none"in this version.relation: (string) The text label representing the semantic relation betweenentity1andentity2. Examples include"Cause-Effect(e1,e2)","Entity-Destination(e1,e2)","Product-Producer(e2,e1)","Other". Note the directionality indicated in the label.text: (string) The full sentence text with the special entity marker tags (<e1>,</e1>,<e2>,</e2>) removed.
How to use
You can load the dataset using the Hugging Face datasets library:
from datasets import load_dataset
# Replace with the actual path on the Hugging Face Hub
dataset_name = "Despina/semeval2010_task8"
dataset = load_dataset(dataset_name)
# Access splits and features
print(dataset['train'][0])
# Expected Output (example):
# {
# 'entity1': 'television',
# 'entity2': 'programmes',
# 'entity1Type': 'none',
# 'entity2Type': 'none',
# 'relation': 'Product-Producer(e2,e1)',
# 'text': 'Most programmes have commercial breaks, but the quality of the television programmes depends on the channel.'
# }
Additional Information
Dataset Curators
This transformed version was generated based on user request using the Hugging Face datasets library. The original dataset was curated by the SemEval-2010 Task 8 organizers.
Licensing Information
The licensing terms for this transformed dataset are inherited from the original SemEvalWorkshop/sem_eval_2010_task_8 dataset. Please refer to the original dataset card or SemEval guidelines for specific licensing information (e.g., CC BY-SA or similar, but confirmation is recommended).
Citation Information
If you use this dataset in your work, please cite the original SemEval 2010 Task 8 paper, as also our paper:
@inproceedings{hendrickx-etal-2010-semeval,
title = "{S}em{E}val-2010 Task 8: Multi-Way Classification of Semantic Relations between Pairs of Nominals",
author = "Hendrickx, Iris and
Kim, Su Nam and
Kozareva, Zornitsa and
Nakov, Preslav and
{'O} S{'e}aghdha, Diarmuid and
Pad{'o}, Sebastian and
Pennacchiotti, Marco and
Romano, Lorenza and
Szpakowicz, Stan",
booktitle = "Proceedings of the 5th International Workshop on Semantic Evaluation",
month = jul,
year = "2010",
address = "Uppsala, Sweden",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/S10-1006",
pages = "33--38",
}
@article{christou2026subbillion,
title = {Sub-Billion, Super-Frontier: Small Language Models Rival
Zero-Shot Frontier LLMs on General and Literary Relation Extraction},
author = {Christou, Despina and Tsoumakas, Grigorios},
journal = {arXiv preprint arXiv:2606.22606},
year = {2026},
url = {https://arxiv.org/abs/2606.22606}
}
Contributions
Thanks to the original SemEval 2010 Task 8 organizers and contributors, and the Hugging Face team for hosting the original dataset.