Relation extraction for graph‐based representation of Hadith Isnad

Abstract

Relation extraction techniques allow the identification of relationships embedded within unstructured text, enabling the transformation of textual data into graph representations for further computational analysis. This thesis applies transformer‐based models to extract narrator relations from hadith texts. Hadith refers to the recorded sayings and actions of the Prophet Muhammad (PBUH), transmitted through chains of narrators known as Isnad. This structure makes hadith particularly well‐suited for relation extraction tasks. Representing isnad as a graph allows for a structured visualization of transmission networks and supports graph‐based analysis for studying hadith authenticity and narrator influence. Despite this potential, relation extraction remains largely unexplored in the context of Hadith literature. To address this gap, this work presents an approach for automatic isnad parsing that converts hadith text into graph‐based representations of narration chains. This thesis tackles two main challenges: the absence of large‐scale structured datasets that reflect the complexity of isnad, and the lack of fine‐tuned relation extraction models tailored to hadith text. A structured dataset was curated from an authoritative digital source, comprising over 270,000 hadith records with annotated narrator mentions and structured narration chains. Using this dataset, two relation extraction approaches were explored: joint learning with an encoder‐based model to identify narrator entities and classify direct narration links, and an encoder‐decoder approach that generates complete chains of narration in an end‐to‐end fashion. Results show that the encoder‐based model is well‐suited for detecting direct links in simpler chains, while the encoder‐decoder models better capture full narration structures, preserving chain integrity and enabling graph construction. The findings from this thesis highlight the potential of relation extraction techniques for advancing computational applications in hadith studies, particularly in isnad analysis. Future work may extend these methods by integrating narrator entity linking and extracting additional properties from hadith texts, contributing to more expressive graph representations of hadith.