From Networks to Named Entities and Back Again

Exploring Classical Arabic Isnad Networks

Abstract

This paper explores new methods for disambiguating the identity of individuals in classical Arabic citations (isnāds) using a network-based approach. After training a model to extract name mentions from classical Arabic, we embed these mentions in vector space using fine-tuned BERT representations and use community detection to infer clusters of coreferent mentions. The best-performing clustering approach reduces error on the CoNLL metric by 30%. Then, as a case study, we examine the problem of determining the number of direct transmitters to Ibn ʿAsākir (d. 1176) in a set of isnāds taken from the 12th century historical text Taʾrīkh Madīnat Dimashq (TMD, History of Damascus), using our method to replicate human judgement.