Machine Learning and Rule Models Decode Historical Hadith Narrator Reliability

Computational researchers have demonstrated that machine learning architectures and expert-encoded rule systems can effectively assess the reliability of historical Hadith narrators by modeling the classical Islamic evaluative discipline known as Al-Jurh and Al-Ta’dil. Meanwhile, three distinct machine learning approaches—Support Vector Machines, Naive Bayes, and Convolutional Neural Networks—converged on an identical test classification accuracy of 88.89 percent, while a Random Forest ensemble achieved 83.33 percent. The findings illustrate how traditional qualitative assessments of transmitter credibility can be mapped systematically into computational feature spaces. By analyzing large-scale biographical data, these algorithmic models provide objective support for textual verification workflows, uncovering complex evaluative patterns while highlighting the complementary strengths of domain-specific human expertise and scalable automated text processing in religious historical scholarship.

In Islamic scholarship, authenticating prophetic traditions requires meticulous scrutiny of both the chain of transmission, known as the Isnad, and the underlying textual narrative, termed the Matn. To distinguish authentic reports from weak or fabricated accounts, classical scholars developed the rigorous science of Al-Jurh and Al-Ta’dil, which centers on the systematic biographical evaluation of every individual narrator in an Isnad. Within this scholarly tradition, Al-Jurh refers to identifying flaws or weaknesses in a narrator that call reliability into question, while Al-Ta'dil refers to affirming a narrator's positive qualities and trustworthiness. Hadith scholars carefully documented the lives of thousands of narrators, including their merits and demerits, based on historical records, contemporary testimony, and their narrations. However, conducting manual cross-examinations of these vast classical volumes is intensely labor-intensive, prompting contemporary researchers to investigate whether computational algorithms can reliably replicate scholarly assessments and accelerate biographical authentication without compromising domain standards.

To build a rigorous experimental pipeline, the research team constructed a structured dataset extracted directly from three authoritative classical dictionaries of narrator criticism: Mizan al-I’tidal by al-Dhahabi, Al-Jurh and Al-Ta’dil by Ibn Abi Hatim, and Al-Jurh and Al-Ta’dil by Yahya ibn Ma’in. The investigators systematically extracted detailed biographical narratives and qualitative assessments for each narrator, recording areas of scholastic consensus as well as notable historical disagreements among the three critics. Analysis of source distributions revealed distinct methodological tendencies among the historical scholars; for example, the laudatory descriptor “truthful” appeared 84 times in Ibn Abi Hatim’s biographical entries, 77 times in al-Dhahabi’s compendium, and 65 times in Yahya ibn Ma’in’s records. This distribution reflects Ibn Abi Hatim’s comparatively lenient evaluative stance compared to the more critical posture maintained by al-Dhahabi. Furthermore, entry lengths varied widely across the corpus, displaying a primary modal concentration between 50 and 100 words and a secondary peak at 140 words, reflecting substantial differences in documentation depth across narrators.

Before inputting the unstructured Arabic prose into machine learning classifiers, the researchers implemented a rigorous multi-stage text preprocessing sequence designed to clean, normalize, and extract informative features from the classical texts. The preprocessing protocol removed extraneous symbols, punctuation marks, and typographical inconsistencies while standardizing historical names, dates, and specialized theological terminology across all entries. Prior to text preprocessing, the raw biographical corpus contained a total of 12,496 individual tokens, with an average length of 105.01 tokens per narrator profile. Following tokenization, stop word filtering, and Arabic root stemming, the total vocabulary count fell to 9,916 tokens, achieving a 20.65 percent reduction in dimensionality that eliminated noise while preserving essential semantic content. Token distribution analysis showed that document lengths, which previously ranged erratically from 50 to over 300 words, stabilized into a normalized clustering centered around 250 words. The top 20 stemmed words exhibited a classic Zipfian distribution, where the most frequent evaluative root appeared more than 600 times, followed by secondary roots appearing 450 to 500 times.

The standout performer in the experimental evaluation was an expert-derived rule-based scoring system constructed strictly from classical methodology literature rather than empirical dataset training. Drawing from canonical treatises including Muqaddimah Ibn al-Salah, Al-Nukat ‘ala Ibn al-Salah by Ibn Hajar al-‘Asqalani, and Al-Raf’a wa al-Takmil by Abd al-Hayy al-Lakhnawi, the team cataloged 47 positive and 39 negative evaluative patterns. Each term received a hierarchical polarity and numerical weight reflecting traditional grades, spanning from plus five for exceptional accolades to minus five for severe repudiations. In addition, evaluating scholar-critics were assigned numerical authority weights reflecting their standing in Hadith criticism based on consensus rankings recorded in classical methodology literature. In an illustrative demonstration analyzing the narrator Bakr bin Khunais, the system combined a mildly positive appraisal from Ali ibn al-Madini with an emphatic critique from Yahya ibn Ma’in, yielding a composite weighted score of minus 1.63 that precisely confirmed the narrator’s documented status of notable weakness.

To confirm that the rule-based model achieved high classification accuracy through genuine predictive inference rather than circular evaluation, the researchers enforced strict procedural independence during ground-truth label generation. Two independent postdoctoral scholars specializing in Hadith sciences reviewed the complete biographical narratives from the classical sources, assigning six-tier reliability labels based on comprehensive contextual reading rather than mechanical keyword matching. Inter-annotator agreement yielded a high Cohen’s kappa coefficient of 0.81, confirming substantial consensus, while remaining discrepancies were arbitrated by a third domain expert. To verify conceptual separation, the authors conducted a lexical perturbation test on 20 held-out narrator entries, replacing classical Hadith terms with neutral Arabic synonyms; the rule-based accuracy immediately plummeted from 95.2 percent to 41.7 percent, demonstrating reliance on genuine domain signals.

In parallel with the rule-based engine, the study implemented four distinct machine learning and deep learning algorithms operating on partitioned datasets split into 70 percent training, 15 percent validation, and 15 percent test sets. A probabilistic Naive Bayes classifier established a standard baseline using term frequency-inverse document frequency vectors to estimate class-conditional probabilities for Arabic evaluation terms. An ensemble Random Forest model incorporated 100 decision trees constructed via bootstrap aggregating, utilizing Gini impurity and information gain to split biographical features and calculate feature importance metrics. A linear Support Vector Machine mapped 1,000-dimensional TF-IDF vectors into an optimal separating hyperplane for multi-class classification across reliability categories. Finally, a Convolutional Neural Network processed sequential text inputs padded to 100 tokens through 128-dimensional embedding layers, rectified linear unit convolutions, max-pooling operations, and dense layers, optimizing classification distributions via categorical cross-entropy loss functions.

Experimental evaluation on the test set revealed remarkable performance consistency across the diverse machine learning architectures despite their fundamentally contrasting mathematical paradigms. Naive Bayes, Support Vector Machines, and the Convolutional Neural Network each achieved identical overall classification test accuracies of 88.89 percent, closely matching their shared validation accuracy of 89.89 percent. The Random Forest classifier demonstrated slightly lower performance under identical experimental conditions, reaching an 83.33 percent test accuracy and an 88.89 percent validation accuracy. Detailed metric evaluations demonstrated identical balanced performance across the three top algorithms, which each recorded a weighted precision of 79.48 percent, a recall of 88.89 percent, and a harmonic F1 score of 83.77 percent. The Random Forest model achieved a precision of 76.09 percent, a recall of 83.33 percent, and an F1 score of 78.49 percent. This close alignment across precision and recall confirmed that the statistical and neural models avoided systematic classification bias across narrator categories.

Analysis of model certainty distributions provided critical insights for designing intelligent human-in-the-loop quality control mechanisms in computational Hadith studies. The Convolutional Neural Network exhibited the highest mean prediction confidence at 96.00 percent, clustering densely in the extreme confidence bracket above 0.95, which researchers noted may indicate slight overconfidence typical of deep networks trained on specialized domain corpora. Conversely, the Support Vector Machine displayed a well-calibrated bimodal confidence distribution averaging 81.71 percent, showing distinct peaks in high-confidence (0.85 to 0.95) and medium-confidence (0.75 to 0.85) intervals. This bimodal behavior naturally separates unambiguous narrator profiles suitable for automated processing from ambiguous records requiring scholarly review. Random Forest and Naive Bayes produced lower mean confidences of 81.11 percent and 79.89 percent, respectively. Confusion matrices further revealed that all models achieved perfect accuracy when identifying unquestionably authentic narrators, while errors occurred exclusively at subtle boundaries between adjacent reliability categories, reflecting historical scholarly debates rather than systemic classification breakdowns.

While the investigation highlights the transformative promise of computational tools in Islamic studies, the researchers noted crucial limitations regarding dataset scope and algorithmic generalizability. Because the dataset was derived from three specific biographical works, the corpus remains modest compared to modern natural language processing benchmarks, which contributed to a performance gap between the CNN’s near-perfect training convergence and its 89 percent validation plateau. Furthermore, while the rule-based approach achieved superior accuracy on curated canonical literature, data-driven machine learning models offer superior practical scalability and resilience when handling linguistically varied, unstandardized, or dialectally diverse biographical texts without demanding continuous manual rule engineering. Ultimately, the authors advocate for a collaborative framework that combines algorithmic efficiency with traditional scholarship, deploying automated models to rapidly verify high-confidence classifications above 90 percent certainty while routing contested boundary cases below 80 percent confidence to human experts for authoritative historical arbitration.

Subject of Research: Automated evaluation of Hadith narrator reliability using machine learning and rule-based models.

Article Title: Automated evaluation of hadith narrator reliability using machine learning and rule-based models based on Al-Jurh and Al-Ta’dil literature

Article References: Wahba, M., Youness, H. A., & Abdelaal, H. M. (2026). Automated evaluation of hadith narrator reliability using machine learning and rule-based models based on Al-Jurh and Al-Ta’dil literature. Discover Artificial Intelligence, 6(1), Article 1832. https://doi.org/10.1007/s44163-026-02247-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02247-3

Keywords: Hadith Authentication, Machine Learning, Al-Jurh wa Al-Ta'dil, Natural Language Processing, Support Vector Machines, Convolutional Neural Networks, Computational Humanities, Text Classification, Automated, evaluation, hadith, narrator