Sangīn

The files in this repository provide for the training and evaluation of a machine learning model that is meant to detect the meter of a hemistich of classical Persian poetry. A recent evaluation run yielded the following results:

Accuracy: 0.9801
F1: 0.9792
Precision: 0.9789
Recall: 0.9801
Loss: 0.0996

Base model

I chose XLM-RoBERTa from Facebook AI. (Is there a better, more recently developed option? If so, I would be happy to switch to it.)

Training data

The data used here comes from Ganjoor. So far, I have added to the dataset the following works, representing a total of 277,248 unique hemistichs:

• The complete ghazals of Ṣāʾib Tabrīzī
• The complete ghazals of Ḥāfiẓ
• The complete ghazals of Saʿdī
• All the ghazals in the Dīvān-i Shams of Rūmī (excepting some with obscure meters)
• The first daftar of the Maṡnavī of Rūmī
• A few thousand lines of the Shāhnāma of Firdawsī
• Four of the poems in the Khamsa of Niẓāmī: Laylī u Majnūn, Khusraw va Shīrīn, the Haft paykar, and the Makhzan al-asrār

More should still be added, but this is a start. The model is already quite good at detecting any of the common meters. Some meters are rare enough that they almost never appear in classical Persian poetry, let alone in this training set.

Plans

• Get the model properly versioned and published
• Add to the training data and adjust training parameters to improve performance
• Present this work at a conference or workshop or similar (if anyone has suggestions...)
• Deploy an inference server and a web front end, i.e., a web app where a user could paste one or more hemistichs from a given poem and have the meter detected (see consensus.py for an idea of how this would work, looking for a consensus of high-confidence meter predictions)