Evaluating the Efficacy of the AraBERT AI Model in Generating Arabic Subject Headings

This study evaluates whether a fine-tuned AraBERT workflow can recommend authorized Arabic subject headings as a controlled-vocabulary shortlist for assigned indexing under ISO 5963. AraBERT is an Arabic-language version of BERT, a transformer model trained to represent the meaning of Arabic text. The recommendation space was Khalifa’s Standard Arabic Subject Headings List (4,696 terms), using 45,000 single-label Arabic book records for training. Gold@K denotes the percentage of records for which the cataloger-assigned correct heading appears within the top K recommendations. On 100 held-out records, the system achieved Gold@1 = 44%, Gold@5 = 84%, and Gold@6 = 86%, supporting cataloger-in-the-loop use.