Sefika's picture
Update README.md
1c639a2 verified
|
Raw
History Blame
4.15 kB
metadata
configs:
  - config_name: default
    data_files:
      - split: test
        path: fewrel_tekgen_inverse.json
license: mit
task_categories:
  - text-classification
  - text-generation
language:
  - en
pretty_name: Reversing Arrows Inverse Relations Benchmark
size_categories:
  - 1K<n<10K
tags:
  - relation-extraction
  - nlp
  - dataset-robustness
  - inverse-relations
  - fewrel
  - tekgen

Reversing Arrows: A Benchmark Dataset for Inverse Relation Directionality in LLMs

This dataset evaluates the robustness of relation extraction (RE) models with a focus on inverse relations and entity substitution. It combines examples from FewRel and TekGen to measure whether models correctly capture relation semantics independent of relation direction and entity identity.

The dataset includes MLCommons Croissant metadata for improved interoperability with machine learning tools and data catalogs.


Dataset Summary

The benchmark investigates whether language models can recognize the same semantic relationship when:

  • the relation direction is reversed (e.g., Mother ↔ Child),
  • entity mentions are replaced with synthetic placeholders,
  • memorized world knowledge is unavailable.

The current benchmark release includes a single combined file for evaluation:

Combined FewRel + TekGen

  • File: data/ablation-tekgen/combined_fewrel_tekgen_inverse.json
  • Records: 5131 (3401 FewRel + 1730 TekGen)
  • Concatenated benchmark file covering both sources.

The repository also provides source files used to build the combined file:

Source Files

  • FewRel source: data/ablation-tekgen/original_fewrel_inverse.json
  • TekGen source: data/ablation-tekgen/original_tekgen_inverse.json

Additionally, the dataset includes:

Relation Metadata

  • File: data/ablation-tekgen/tekgen_relations.json
  • Human-readable relation and inverse-relation definitions for TekGen relations.

Croissant Metadata

  • File: data/ablation-tekgen/inverse_relations_croissant.json
  • MLCommons Croissant description of available resources and fields.

--

Dataset Structure

Core evaluation file:

File Description
fewrel_tekgen_inverse.json` Combined benchmark examples from FewRel and TekGen.

Features

Because the dataset combines two sources, fields vary slightly by source. Common fields are:

Field Description
head_to_tail Forward relation label (PID for FewRel or relation string for TekGen).
tail_to_head Inverse relation label.
head Head entity.
tail Tail entity.

FewRel-specific fields:

Field Description
tokens Tokenized sentence.
relation_prop_wiki Wikidata relation URIs connected to the entity pair.

TekGen-specific fields:

Field Description
id Unique example identifier.
sent Full sentence string.
artificial_data Mapping from original to substituted entities (only in artificial TekGen file).

Usage

Load the combined benchmark file with Python:

import json

with open("data/ablation-tekgen/combined_fewrel_tekgen_inverse.json", "r", encoding="utf-8") as f:
  dataset = json.load(f)

Potential applications include:

  • Relation Extraction
  • Text Classification
  • Robustness Evaluation
  • Knowledge Graph Completion
  • Directionality Analysis
  • Entity Generalization Studies

Repository

Scripts used to generate the inverse relations and synthetic entity substitutions are available at:

GitHub: https://github.com/sefeoglu/inverserelations


Citation

If you use this dataset, please cite both the original FewRel paper and this dataset.

@misc{efeoglu2025inversefewrel,
  author       = {Sefika Efeoglu and Adrian Paschke},
  title        = {Reversing Arrows: A Benchmark Dataset for Inverse Relation Directionality in LLMs},
  year         = {2025},
  doi          = {10.57967/hf/8462},
  publisher    = {Hugging Face},
  howpublished = {\url{https://github.com/sefeoglu/inverserelations}}
}