Despina/conll04

Processed CoNLL-2004 Relation Extraction Dataset

Dataset Description

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 processed version of the DFKI-SLT/conll04 dataset, tailored for relation extraction tasks. The original dataset, based on the CoNLL-2004 shared task data but adapted for relation extraction, presents sentences with lists of entities and relations. This version transforms the data into a "flattened" format where each row represents a single relationship instance.

For every relation identified in the original dataset between a head entity and a tail entity, a new record is created containing:

• The text spans of the head and tail entities.
• The types of the head and tail entities.
• The type of the relation itself.
• The full text of the sentence containing the relation.
• The original sentence identifier (orig_id).

This format can be directly useful for training models that predict the relation type given the sentence context and the identified entities.

Source Data

The data originates from the DFKI-SLT/conll04 dataset available on the Hugging Face Hub. That dataset itself appears to be based on data from the CoNLL-2004 shared task, adapted for relation extraction research (commonly associated with work by Roth and Yih, e.g., 2004).

Data Fields

The dataset contains the following fields:

• orig_id (string): Original identifier for the sentence from the source dataset.
• text (string): The full text of the sentence from which the relation was extracted.
• entity1 (string): The text span of the head entity in the relation.
• entity2 (string): The text span of the tail entity in the relation.
• entity1Type (string): The type/category of the head entity (e.g., Peop, Org, Loc, Other).
• entity2Type (string): The type/category of the tail entity (e.g., Peop, Org, Loc, Other).
• relation (string): The type of relation between entity1 (head) and entity2 (tail) (e.g., Work_For, Located_In, Kill).

Data Splits

The dataset provides the following splits:

• train
• validation
• test

Data Creation

This dataset version was generated using a Python script that processes the DFKI-SLT/conll04 dataset. The script iterates through each example (sentence) in the train and test splits. For each relation listed in an example, it:

1. Retrieves the head and tail entity indices.
2. Looks up the corresponding entity information (type, start/end token indices) from the entities list.
3. Extracts the entity text spans by slicing the sentence's token list.
4. Joins the sentence's token list to create the full sentence text.
5. Retrieves the original sentence ID.
6. Combines these pieces of information into a single record for the new dataset.

Intended Use

This dataset is primarily intended for training and evaluating supervised models for Relation Extraction. Given a sentence and two entities within it, the goal is typically to predict the semantic relation between them. The flattened format is convenient for this purpose.

Limitations

• Inherited Limitations: The dataset inherits any potential limitations or biases present in the original CoNLL-2004 data and the DFKI-SLT/conll04 adaptation, such as domain specificity (likely newswire) and potential annotation inconsistencies.
• Data Duplication: The full sentence text (text) and original ID (orig_id) are duplicated for every relation instance extracted from the same sentence.

Citation Information

If you use this dataset, please cite the work associated with the original CoNLL-2004 relation extraction data formulation, as also our work:

@inproceedings{roth-yih-2004-linear,
    title = "A Linear Programming Formulation for Global Inference in Natural Language Tasks",
    author = "Roth, Dan  and
      Yih, Wen-tau",
    booktitle = "Proceedings of the Eighth Conference on Computational Natural Language Learning ({C}o{NLL}-2004)",
    month = may,
    year = "2004",
    address = "Boston, Massachusetts, USA",
    publisher = "Association for Computational Linguistics",
    url = "[https://aclanthology.org/W04-2401](https://aclanthology.org/W04-2401)",
    pages = "1--8",
}

@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}
}