Fine-Tuning BERT for Persian Poet Identification

In this experiment, we have attempted the task of poet identification for four prominent Persian poets (i.e., Hafez, Omar Khayyam, Rumi, and Saadi Shirazi) by fine-tuning the BERT language representation model on a dataset of hemistichs. Among the challenges of this task was the imbalanced distribution of the hemistichs, with one class containing more than ۵۲۰۰۰ hemistichs while another class contained less than ۱۳۰۰ hemistichs. Moreover, the short length of the hemistichs made this task more challenging than poet identification using a whole poem or even a verse. It was also demonstrated that the diction used by the poets was similar to some degree, which further added to the challenge of the task at hand. The model attained a Matthews correlation coefficient of ۰.۶۴۶ on the test set, and the effectiveness of transfer learning in processing works of literature was demonstrated even in case of unsubstantial data, an imbalanced dataset, and similar diction.

کلیدواژه ها:

author identification ، digital humanities ، imbalanced data ، poet identification ، computational linguistics ، bert