NLI-RE: Arabic Relation Extraction as Natural Language Inference

NLI-RE reformulates relation extraction (RE) as a Natural Language Inference (NLI) task. Given an input sentence containing two named entity mentions, the sentence is treated as the premise and a hypothesis is constructed by verbalizing a candidate relation between the two entities using a predefined Arabic template. A transformer encoder then predicts whether the premise entails the hypothesis, indicating the presence of that relation between the entity pair.

---

Table of Contents

• Framework Overview
• Project Structure
• Installation
• Dataset
• Configuration
• Usage
• Citation
• License

---

Framework Overview

NLI-RE consists of three steps:

1. Template Construction

For each of the 40 relation types, a relation-aware Arabic template maps a candidate relation between subject entity e1 and object entity e2 into a hypothesis string. For example:

Relation	Subject	Object	Hypothesis
Location.located_in	قبة الصخرة	مدينة القدس	قبة الصخرة تقع في مدينة القدس
Personal.has_occupation	حبيب بولس	مدير مال عكا	حبيب بولس يعمل كـ / مهنته مدير مال عكا
Affiliation.member_of	أطباء بلا حدود	الأمم المتحدة	أطباء بلا حدود عضو في الأمم المتحدة

2. Sentence Encoding

A transformer encoder (UBC-NLP/ARBERTv2) processes the concatenated premise-hypothesis pair:

[CLS] sentence [SEP] hypothesis

3. Relation Inference

The encoded representation is passed through a fully connected layer producing a binary prediction:

• True → the relation holds between the entity pair in the sentence
• False → the relation does not hold

Training Objective

The model is trained with a combined loss to handle class imbalance between positive (relation present) and negative (relation absent) instances:

Loss = L_WCE + L_NCE

• L_WCE — Weighted Cross-Entropy with per-class weights [w_n, w_p], where w_p > w_n penalises missed positive instances more heavily.
• L_NCE — Noise Contrastive Estimation with temperature τ to improve separation between positive and negative instances.

---

Project Structure

arabic-nli-classifier/
│
├── configs/
│   └── config.yaml                  # Hyperparameters and paths
│
├── data/
│   ├── nli/                         # NLI pairs included in the repository
│   │   ├── train.jsonl              # 280 premise-hypothesis pairs
│   │   ├── val.jsonl                #  60 pairs
│   │   └── test.jsonl               #  60 pairs
│   └── raw/                         # Place full dataset here (not committed)
│       └── train.jsonl
│
├── src/
│   ├── data/
│   │   ├── templates.py             # 40 Arabic relation verbalization templates
│   │   ├── nli_generator.py         # Generates [CLS] s [SEP] h pairs from RE records
│   │   ├── dataset.py               # PyTorch Dataset for NLI sentence pairs
│   │   └── loader.py                # Loads NLI JSONL splits
│   │
│   ├── models/
│   │   └── trainer.py               # ContrastiveTrainer: Loss = L_WCE + L_NCE
│   │
│   ├── training/
│   │   └── cross_val.py             # Stratified K-Fold fine-tuning loop
│   │
│   ├── evaluation/
│   │   └── metrics.py               # Micro-F1, ensemble inference, reporting
│   │
│   └── utils/
│       └── helpers.py               # Seed control, label maps, config loader
│
├── scripts/
│   ├── prepare_data.py              # Prepare NLI pairs from raw RE records
│   ├── train.py                     # Run K-Fold training
│   └── predict.py                   # Run ensemble inference
│
├── requirements.txt
├── setup.py
└── README.md

---

Installation

git clone https://github.com/alaa-aljabari/ArabicNLI-RE.git
cd ArabicNLI-RE

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

pip install -r requirements.txt

---

Dataset

Included sample

The repository includes a sample of 400 NLI pairs in data/nli/:

File	Pairs	True	False
train.jsonl	280	138	142
val.jsonl	60	38	22
test.jsonl	60	30	30

Each record contains:

{
  "nli_sentence": "[CLS] رسالة من مدير مال عكا حبيب بولس ... [SEP] حبيب بولس يعمل كـ / مهنته مدير مال عكا",
  "Label": "True"
}

Full dataset (WojoodRelations)

The full WojoodRelations dataset is available via registration:

📥 Download WojoodRelations

Tools, datasets, and models are also available through the online demo:

🔗 Arabic Relation Extraction Tools, Datasets and Models

Using your own data

Place your raw RE records at data/raw/train.jsonl, then run:

python scripts/prepare_data.py --config configs/config.yaml

# Custom sample size:
python scripts/prepare_data.py --n_positive 500 --n_negative 500

Raw records must contain the fields: sentence, subject, object, relation.

---

Configuration

model:
  name: "UBC-NLP/ARBERTv2"   # Sentence encoder
  max_len: 128                 # Max tokens for [CLS] s [SEP] h
  num_folds: 2                 # K in Stratified K-Fold

loss:
  tau: 1.0                     # Temperature τ for L_NCE
  class_weights: [0.2, 1.0]   # [w_n, w_p] for L_WCE

data:
  train_path: "data/nli/train.jsonl"
  dev_path:   "data/nli/val.jsonl"
  test_path:  "data/nli/test.jsonl"

---

Usage

Run on the included sample (no setup needed)

# Train
python scripts/train.py --config configs/config.yaml

# Predict
python scripts/predict.py --config configs/config.yaml

Run on the full dataset

# Step 0 — generate NLI pairs from raw RE records
python scripts/prepare_data.py --config configs/config.yaml

# Step 1 — train
python scripts/train.py --config configs/config.yaml

# Step 2 — predict
python scripts/predict.py --config configs/config.yaml

Google Colab

A ready-to-run Colab notebook is available. Click the badge below to open it directly:

---

Citation

If you use this code or the dataset in your research, please cite:

@inproceedings{aljabari-etal-2025-wojoodrelations,
    title = "$\mathrm{Wojood^{Relations}}$: {A}rabic Relation Extraction Corpus and Modeling",
    author = "Aljabari, Alaa  and
      Khalilia, Mohammed  and
      Jarrar, Mustafa",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.1741/",
    doi = "10.18653/v1/2025.emnlp-main.1741",
    pages = "34342--34360",
    ISBN = "979-8-89176-332-6",
}

@inproceedings{aljabari-etal-2025-wojoodontology,
    title = "{W}ojood{O}ntology: Ontology-Driven {LLM} Prompting for Unified Information Extraction Tasks",
    author = "Aljabari, Alaa  and
      Hamad, Nagham  and
      Khalilia, Mohammed  and
      Jarrar, Mustafa",
    booktitle = "Proceedings of The Third Arabic Natural Language Processing Conference",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.arabicnlp-main.14/",
    doi = "10.18653/v1/2025.arabicnlp-main.14",
    pages = "179--193",
    ISBN = "979-8-89176-352-4",
}

@inproceedings{aljabari-etal-2024-event,
    title = "Event-Arguments Extraction Corpus and Modeling using {BERT} for {A}rabic",
    author = "Aljabari, Alaa  and
      Duaibes, Lina  and
      Jarrar, Mustafa  and
      Khalilia, Mohammed",
    booktitle = "Proceedings of the Second Arabic Natural Language Processing Conference",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.arabicnlp-1.26/",
    doi = "10.18653/v1/2024.arabicnlp-1.26",
    pages = "309--319",
}

---

License

MIT