Fine-tuned transformer models for automatic isnad parsing: extracting narrator mentions and narration relations (who narrated to whom) from hadith isnad text. Use the output to build graph representations of transmission chains for authenticity analysis, visualization, and retrieval.
The isnad is the chain of narrators at the start of a hadith; it records who transmitted the report across generations. Isnad parsing means:
Narrator recognition — identifying narrator mentions (and optionally narration words) in the isnad text.
Relation extraction — extracting direct narration links as (head, tail) pairs: the head is the receiver of the hadith, the tail is the transmitter.
When a hadith has multiple chains (branching or parallel paths), we can also predict a chain identifier per link, giving (head, tail, chain_number) triplets so that each transmission path stays distinct in the graph.
Chain-blind: output is duplets (head, tail) only; no chain IDs. Suitable for a single merged graph of all links (no validation on paths).
Chain-aware: output is triplets (head, tail, chain_number); suitable for reconstructing separate paths and building property graphs with chain IDs on edges.
Example: The following isnad encodes three distinct transmission chains that share some narrators but branch into different paths. Preserving a chain identifier per relation is necessary to avoid invalid narration paths when building the graph.
Isnad (Arabic and translation):
حدثنا مسدد ، حدثنا يحيى ، عن سفيان قال: حدثني منصور وسليمان ، عن أبي وائل ، عن أبي ميسرة ، عن عبد الله . قال: وحدثني واصل ، عن أبي وائل ، عن عبد الله - رضي الله عنه...
Musaddad narrated to us, Yahya narrated to us, from Sufyan, who said: Mansur and Sulaiman narrated to me, from Abu Wa'il, from Abu Maysarah, from Abdullah. He said: Wasil also narrated to me, from Abu Wa'il, from Abdullah – may Allah be pleased with him...
This hadith isnad has three narration chains. They share the same opening (Musaddad ← Yahya ← Sufyan) but then branch:
Chain 1 goes through Wasil: Sufyan ← Wasil ← Abu Wa'il ← Abdullah.
Chain 2 goes through Sulaiman: Sufyan ← Sulaiman ← Abu Wa'il ← Abu Maysarah ← Abdullah.
Chain 3 goes through Mansur: Sufyan ← Mansur ← Abu Wa'il ← Abu Maysarah ← Abdullah.
Without chain numbers, merging all (head, tail) links into one graph would mix these paths and create invalid connections. The figures below show each chain separately, then the merged graph with chain numbers on the edges.
Chain 1 (Wasil path):
Graph representation of chain 1
Chain 2 (Sulaiman path):
Graph representation of chain 2
Chain 3 (Mansur path):
Graph representation of chain 3
Merged graph (all three chains, with chain numbers on edges):
Complete hadith isnad as narration graph
Example: model input and output
The models take raw isnad text and generate a linearized sequence of relation tuples. Below: one input and its chain-blind and chain-aware outputs.
Input (Arabic isnad as in source text). English translation for demonstration:
حدثنا يوسف القاضي ، ثنا عبد الواحد بن غياث ، ثنا حماد بن سلمة ، عن أيوب وعبيد الله بن عمر ، عن نافع ، عن ابن عمر قال : قال زيد بن ثابت...
Yusuf Al-Qadi narrated to us, saying: Abdul Wahid bin Ghayath narrated to us, saying: Hammad bin Salamah narrated to us, from Ayyub and Ubaidullah bin Umar, from Nafi', from Ibn Umar, who said: Zaid bin Thabit said...
Chain-blind output (duplets only; no chain IDs). Each relation is <head>receiver<tail>transmitter; the sequence ends with #hadith.:
<head> يوسف القاضي <tail> عبد الواحد بن غياث
<head> عبد الواحد بن غياث <tail> حماد بن سلمة
<head> حماد بن سلمة <tail> عبيد الله بن عمر
<head> حماد بن سلمة <tail> أيوب
<head> عبيد الله بن عمر <tail> نافع
<head> أيوب <tail> نافع
<head> نافع <tail> ابن عمر
<head> ابن عمر <tail> زيد بن ثابت #hadith
Chain-aware output (triplets with chain identifier). Each relation adds <chain_number>k so that the two transmission paths stay distinct.:
<head> يوسف القاضي <tail> عبد الواحد بن غياث <chain_number> 1
<head> عبد الواحد بن غياث <tail> حماد بن سلمة <chain_number> 1
<head> حماد بن سلمة <tail> عبيد الله بن عمر <chain_number> 1
<head> عبيد الله بن عمر <tail> نافع <chain_number> 1
<head> نافع <tail> ابن عمر <chain_number> 1
<head> ابن عمر <tail> زيد بن ثابت <chain_number> 1
<head> يوسف القاضي <tail> عبد الواحد بن غياث <chain_number> 2
<head> عبد الواحد بن غياث <tail> حماد بن سلمة <chain_number> 2
<head> حماد بن سلمة <tail> أيوب <chain_number> 2
<head> أيوب <tail> نافع <chain_number> 2
<head> نافع <tail> ابن عمر <chain_number> 2
<head> ابن عمر <tail> زيد بن ثابت <chain_number> 2
#hadith
Model architectures
Base model
Type
Description
CAMeLBERT-CA
Encoder (BERT)
Pre-trained on Classical Arabic. Joint NER + relation classification (biaffine). Outputs chain-blind duplets only. Best for high-precision direct links.
AraT5
Encoder–decoder (T5)
T5 for Arabic; text-to-text. Can be fine-tuned for chain-blind or chain-aware output. Best for chain-aware triplets and full graph reconstruction.
AraBART
Encoder–decoder (BART)
BART for Arabic; denoising pretraining. Can be fine-tuned for chain-blind or chain-aware output. May show more spelling variation; useful for tasks like narrator normalization.
With-prefix vs without-prefix: Some models were trained with a task prefix (e.g. [isnad]) in the input; others without. Use the variant that matches your intended input format.
Hugging Face models
All models are hosted on the Hugging Face Hub. Choose by task (chain-blind vs chain-aware) and base architecture (CAMeLBERT-CA, AraT5, AraBART).
Source. Training and evaluation data come from the Ifta' Sunnah Platform (King Abdullah bin Abdul Aziz Program for the Prophetic Sunnah), and curated for the experiment at Ifta-Sunnah–Hadith-and-Narrators-Dataset. The platform provides 33 foundational hadith books (Kutub al-Matn), plus 57 additional books with supportive and interpretative content; the hadith dataset used here was extracted from the 33 books, which are organized mainly by hadith content (thematic). In total, 276,347 hadiths were extracted from the platform; after preprocessing (isnad isolation, entity tagging, relation tuples, and filtering), 223,478 hadith instances were used for automatic isnad parsing.
Splits. The dataset was split 70% training, 15% validation, 15% testing. Summary:
Training
Validation
Testing
Hadiths
156,434
33,521
33,523
Duplets (chain-blind relations)
756,714
162,782
162,010
Triplets (chain-aware relations)
877,016
190,049
188,755
Results summary (test set)
Models are evaluated on the held-out test set (33,523 hadith isnad texts). Metrics: exact match for relations; narrator set comparison for narrator recognition. Tables below include with-prefix and no-prefix variants for the encoder–decoder models.
Narrator recognition (predicted vs gold set of narrators)
Model
Precision
Recall
F1
CAMeLBERT-CA
0.98
0.99
0.98
AraT5
0.97
0.97
0.97
AraT5-Prefix
0.97
0.97
0.97
AraBART
0.78
0.94
0.84
AraBART-Prefix
0.78
0.94
0.84
Chain-blind relation extraction (duplets)
Model
Precision
Recall
F1
CAMeLBERT-CA
0.96
0.93
0.95
AraT5
0.96
0.95
0.95
AraT5-Prefix
0.96
0.95
0.95
AraBART
0.66
0.64
0.65
AraBART-Prefix
0.66
0.65
0.65
Chain-aware relation extraction (triplets)
Only the encoder–decoder models produce chain-aware output.
The CAMeLBERT-CA Hub repo provides model.safetensors; the inference script loads that when using --model-id.
How to use
CAMeLBERT-CA (chain-blind): inference script in this repo
The CAMeLBERT-CA model does not output duplets directly from raw text. It needs a pipeline that: (1) runs NER to get narrator spans, (2) builds candidate head–tail duplets, and (3) runs the biaffine classifier on each candidate. This repo provides that pipeline in camelbert_duplets_biaffine_inference.py.
Run inference (loads the model from Hugging Face by default):
bash
1# Built-in example isnad2python camelbert_duplets_biaffine_inference.py --no-cuda
34# Your own isnad text (space-separated words)5python camelbert_duplets_biaffine_inference.py --input "حدثنا أبو معاوية حدثنا الأعمش عن مسلم" --no-cuda
67# From a JSONL file (one JSON object per line with "isnad_sequence" or "isnad_text")8python camelbert_duplets_biaffine_inference.py --input-file examples.jsonl --no-cuda
Output: One line per input in chain-blind form: <head> A <tail> B <head> B <tail> C ... #hadith
To use a local checkpoint (same format as training output): --checkpoint path/to/best_model.pth
Test on random examples from the predictions JSON (install ijson for large files):
These models generate the linearized output directly from raw isnad text:
python
1from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
23model_name ="Jehadoumer/ChainBlind-HadithIsnadParser-AraT5"# or any other model ID above4tokenizer = AutoTokenizer.from_pretrained(model_name)5model = AutoModelForSeq2SeqLM.from_pretrained(model_name)67# Input: raw isnad text (Arabic)8isnad_text ="..."# e.g. "حدثنا فلان، حدثنا فلان، عن فلان..."9inputs = tokenizer(isnad_text, return_tensors="pt", truncation=True, max_length=512)10outputs = model.generate(**inputs, max_new_tokens=512)1112# Decode and parse <head> / <tail> (and optionally <chain_number>) from the output13decoded = tokenizer.decode(outputs[0], skip_special_tokens=False)14# Then extract (head, tail) or (head, tail, chain_number) from decoded.
Output format: chain-blind — sequence of <head> A <tail> B segments ending with #hadith. Chain-aware — each segment also includes <chain_number> k. Parse with regex or the same logic as in the evaluation scripts.
If you use these models or the dataset in your work, please cite the thesis and the original data source as appropriate.
License
This project is released under an open-source license that allows use, modification, and distribution with attribution to the creators. The base models (CAMeLBERT-CA, AraT5, AraBART) follow their respective upstream licenses; the fine-tuned weights are released under the same open-source, attribution-based terms.
