Automatic Romanization of Arabic Bibliographic Records
This addresses the tedious manual Romanization required by international library standards for cataloguers and users without Arabic expertise, representing a novel application but incremental in method.
The paper tackles the problem of automatically Romanizing undiacritized Arabic bibliographic records, achieving 89.3% exact word Romanization accuracy on a blind test set.
International library standards require cataloguers to tediously input Romanization of their catalogue records for the benefit of library users without specific language expertise. In this paper, we present the first reported results on the task of automatic Romanization of undiacritized Arabic bibliographic entries. This complex task requires the modeling of Arabic phonology, morphology, and even semantics. We collected a 2.5M word corpus of parallel Arabic and Romanized bibliographic entries, and benchmarked a number of models that vary in terms of complexity and resource dependence. Our best system reaches 89.3% exact word Romanization on a blind test set. We make our data and code publicly available.