Tarikhi: Arabic Temporal Information Extraction from Arabic Historical Documents

Abstract

Arabic historical books and archival materials contain rich accounts of political, social, and cultural events, yet they remain largely underutilized computationally due to the scarcity of dedicated Arabic information extraction tools. The challenge is amplified in long-form, scanned historical documents, where optical character recognition noise, orthographic variation, and complex narrative structures complicate automatic processing. In this paper, we present Tarikhi, a retrieval-augmented generation framework for structured temporal event extraction from Arabic scanned books. The proposed pipeline integrates high-accuracy optical character recognition, chunking-based processing for long-document handling, Arabic named entity recognition, span refinement, and a retrieval-enhanced attribute extraction module that identifies event dates, locations, and descriptive summaries. Extracted events are consolidated and linked using semantic and temporal similarity measures, and linked through relation classification to construct structured temporal events. Evaluation on a selected part of modern Arabic historical books demonstrates the feasibility of temporal event extraction from long-form Arabic texts, achieving a 75.3% F1-score under dual human verification. Tarikhi represents a step toward scalable temporal knowledge construction for Arabic digital humanities resources.

Resources

Details