| --- |
| license: mit |
|
|
| task_categories: |
| - text-classification |
| - text-generation |
| language: |
| - en |
| pretty_name: Relation Extraction for Inverse Relations FewRel |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - relation-extraction |
| - nlp |
| - dataset-robustness |
| - inverse-relations |
| --- |
| |
| <script type="application/ld+json"> |
| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "cr": "http://mlcommons.org/croissant/", |
| "sc": "https://schema.org/" |
| }, |
| "@type": "sc:Dataset", |
| "name": "InverseRelationsBenchmark", |
| "description": "Benchmark based on FewRel with original entity mentions and synthetic substitutions for robustness testing.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0" |
| } |
| </script> |
| |
| # Reversing Arrows: A Benchmark Dataset for Inverse Relation Directionality in LLMs |
|
|
| This dataset is designed to evaluate the robustness of Relation Extraction (RE) models, with a specific focus on **Inverse Relations** and entity substitution. It is derived from **FewRel**. |
| The croissant metadata is available on [croissant](https://huggingface.co/datasets/Sefika/FewRel_Inverse_Relations/blob/main/metadata.txt) |
| ## Dataset Summary |
|
|
| The primary goal of this dataset is to test whether models can recognize a relationship regardless of the direction or the surface form of the entities. It includes: |
| 1. **Inverse Mapping**: Every sample includes both the forward relation (e.g., *Mother*) and its logical inverse (e.g., *Child*). |
| -- *Original data* : original_fewrel_inverse.json |
| 3. **Synthetics Substitutions**: A version of the dataset where real-world entities are replaced with synthetic names (e.g., replacing "Denmark" with "Emilyfort") to test if models rely on memorized entity knowledge rather than linguistic context. |
| -- *synthetic_fewrel_inverse.json* |
| ### Relation Examples |
| | Forward Relation | Inverse Relation | |
| | :--- | :--- | |
| | **Child** (P40) | **Mother** (P25) | |
| | **Child** (P40) | **Father** (P22) | |
| | **Follows** (P155) | **Followed by** (P156) | |
| | **Has Part** (P527) | **Part of** (P361) | |
| |
| ## Dataset Structure |
|
|
| The dataset is provided in JSON format and split into four main distributions: |
|
|
| * **Original FewRel**: The ground-truth sentences and relations. (*original_fewrel_inverse.json*) |
| * **Synthetic FewRel**: Sentences where entities have been substituted with synthetic placeholders. (*synthetic_fewrel_inverse.json*) |
| * **Relation Metadata**: Supplemental files providing human-readable definitions for Wikidata P-IDs (*fewrel_inverse_relations.json*) |
|
|
| ### Key Fields |
|
|
| * `tokens`: The input text (tokenized or raw string) in FewRel. |
| * `head` / `tail`: The subject and object entities. |
| * `head_to_tail`: The ID/Name of the forward relation. |
| * `tail_to_head`: The ID/Name of the inverse relation. |
| * `artificial_data`: A mapping showing which `Original` entities were changed to which `Synthetic` placeholders. |
|
|
| ## Usage |
|
|
| You can load this dataset for tasks such as: |
| * **Text Classification**: Predicting the relation ID between two entities. |
| * **Text Generation**: Generating a sentence that expresses a specific inverse relation. |
|
|
| ### Repository |
|
|
| For scripts and more technical details on how the artificial data was generated, visit the official repository: |
| [sefeoglu/inverserelations](https://github.com/sefeoglu/inverserelations) |
|
|
| ## Citation |
|
|
| If you use this dataset in your research, please cite the original FewRel and this repository. |
|
|
| ```bibtex |
| |
| @misc{fewrel_inverse_2025, |
| author = {Sefika Efeoglu, and Adrian Paschke}, |
| title = { Reversing Arrows: A Benchmark Dataset for Inverse Relation Directionality in LLMs}, |
| year = {2025}, |
| publisher = {GitHub/HuggingFace}, |
| doi = { 10.57967/hf/8462 }, |
| howpublished = {https://github.com/sefeoglu/inverserelations} |
| } |
| |
| |
